From Scarcity To Abundance

Kommentar verfassen

What does the world look like when we are no longer compute constrained?

Billions of dollars are flowing into AI companies optimized for a world that may not exist in five years.

Right now, AI is massively compute constrained. And in just the past few weeks, the evidence is piling up:

Even the favorite shoe brand of every tech bro in 2017 is looking to get into the compute game…

All of this is real. The scarcity is real. And the companies capitalizing on it are posting incredible numbers and building meaningful businesses. But here’s the thing: every prior technology cycle had a scarce resource at its center, and every time, that scarcity eventually broke. When it did, the value map reshuffled dramatically – and the companies that looked unassailable during the scarcity era often weren’t the long-term winners.

Our view: the venture landscape is dramatically overweighting what is in demand today versus what is going to be in demand for the next ten years.

We’ve Seen This Before

 

Every prior innovation cycle started with a scarce resource that eventually became abundant.

Oil in the 19th century

In the late 1800s, crude was actually plentiful. Wildcatters kept finding more of it. The real bottleneck was refining capacity and distribution, which is why John D. Rockefeller built Standard Oil around those layers rather than drilling. But once refining technology matured and pipeline networks expanded, that bottleneck broke too. And the interesting thing is what came next: the automotive economy, petrochemicals, plastics, commercial aviation. Ford’s Model T only made sense because fuel was getting cheap. The plastics revolution required abundant petroleum feedstocks. These were entire industries that nobody was really thinking about during the scarcity era, and they ended up dwarfing the value of oil extraction and refining combined.

Telecom in the 1990s

The telecom boom of the late ’90s followed a similar arc, with a twist. Bandwidth was the scarce resource, and telecoms raised hundreds of billions to control it. Then the bubble burst; and the bust created the abundance. All that fiber didn’t disappear when Global Crossing and WorldCom went bankrupt. It got bought at pennies on the dollar. What happened next is instructive. Google, YouTube, Netflix, Spotify — none of these businesses were economically viable at 1999 bandwidth prices. They needed cheap bandwidth to exist at all. Meanwhile, the companies that had been valuable specifically because bandwidth was expensive got crushed. RealNetworks, once worth over $30 billion for its streaming compression tech, became irrelevant almost overnight. Why bother with clever compression when you can just send the full stream? CDN technology went from a high-margin standalone business to a feature baked into cloud platforms. Even Salesforce and the broader SaaS model were downstream beneficiaries of cheap, reliable connectivity.

The winners weren’t the ones who owned the scarce resource or built optimization tricks around it. They were the ones who built for the world where it was cheap.

Cloud / SaaS in the 2010s

Cloud and SaaS repeated the pattern one more time. Through the 2010s, the bottleneck was engineering talent and scalable infrastructure. Engineer salaries soared. Companies fought viciously over hiring. A legendary show satirizing Silicon Valley culture became required viewing. SaaS pricing reflected the genuine cost of building and maintaining good software. Then AWS, Azure, and GCP commoditized infrastructure, open source commoditized components, and value migrated again — from horizontal platforms to vertical SaaS with deep domain knowledge, from engineering as the moat to distribution as the moat, from building software to configuring it. Some of the most valuable late-stage SaaS companies weren’t particularly technically impressive. They just had the best go-to-market and the deepest workflow integration. Same story: when the scarce resource got cheap, value moved up the stack toward whoever was closest to the end user and the actual problem being solved.

These prior waves prove that as the scarce resource becomes abundant, value migrates UP THE STACK. Applications, workflows, and things that touch the user accrue value; “optimization” layers that were valuable during scarcity get squeezed.

DrPeering White Paper - Internet Transit Prices - Historical and Projections
 
Bandwidth prices in the 90s and early 2000s got cheap fast

What’s Priced For Scarcity Today?

 

It feels like fundraising and commercialization in the AI market today is heavily skewed towards companies capitalizing on the compute shortages. On the public side, chip companies like Nvidia have been making hay for the past few years, but nearly every company across memory (Sandisk, SK Hynix), semis (TSMC, AMD), and power (Bloom Energy, Vistra) have been seeing record revenues, profits, and a ripping stock. This makes sense — during scarcity, the resource providers always have the best economics. The question is how much of this is structural versus cyclical.

 
When a 35 year old company has what everyone wants

We’re seeing the same dynamic play out across the private markets. Heavy funding, rapid ARR growth, and massive valuations in categories that are fundamentally downstream of expensive compute include:


  • Inference optimization and RL reasoning. Companies like Together AI, Baseten, and Fireworks are building real, fast-growing businesses around making inference faster and cheaper. This is a fantastic category to be in when compute is expensive and generating more intelligence per dollar absolutely matters. On the other end, think about what happened to RealNetworks, or to Akamai’s pricing power once bandwidth got cheap. You don’t need clever compression tricks when you can just send the full stream. When compute gets cheap, you can brute-force a lot of what these techniques achieve — run a bigger model, run multiple passes, throw more inference at the problem and pick the best answer. The techniques won’t disappear, but core pricing will likely commoditize. So the question becomes whether these businesses will have to refactor to maintain durable pricing power in a world with abundant compute.



  • GPU access and compute brokering. CoreWeave is probably the most prominent example, but there’s a whole cohort of GPU cloud companies — Lambda Labs, Crusoe, and others — that have raised significant capital on the back of GPU scarcity. The core (dumbed down) value proposition for most of these companies is “we have allocation.” That is a terrific in a supply-constrained world. The question is how sustainable that moat is when that constraint goes away. What happens when the arb goes away?



  • Model training and tooling. When a single frontier training run costs tens or hundreds of millions of dollars and a failed run is a catastrophe, the willingness to pay for anything that makes that process more reliable and efficient is enormous. That math changes pretty quickly if compute costs drop by an order of magnitude.


None of this means these are bad companies or bad technologies. In fact sometimes its the exact opposite. The pattern from prior cycles isn’t that the scarce-resource companies go to zero — Exxon is still enormous, Akamai still exists, AWS still prints money.

The point is that during scarcity, the market tends to OVERVALUE these layers and UNDERVALUE what comes next. The best returns in the oil era didn’t come from refining. The best returns in the internet era didn’t come from owning fiber. And the best returns in AI might not come from the layers that look most valuable right now.

 

Where Will Value Migrate In The Abundance Era?

 

When compute and infrastructure are no longer the bottleneck, AI goes from supply-constrained to demand-constrained. And in demand-constrained markets, the moats that have always mattered reassert themselves: user attention, distribution, brand, workflow integration, and switching costs.

So what areas do we think will flourish in an era of cheap compute?


  • Vertical applications that own the user relationship. Companies embedded in real workflows with proprietary data accumulated through thousands of customer interactions — legal AI built on real contract negotiations, security platforms with proprietary threat data, healthcare AI woven into clinical decisions. The test: if every model becomes equally capable and cheap tomorrow, does your company still matter? If yes because you own the distribution or you’re too embedded to rip out, you’re on the right side. These companies‘ margins actually expand as compute gets cheaper, which is the opposite of what happens to the optimization layer.



  • Physical AI, robotics, and space. When compute is cheap, the constraint shifts from „can we run the model“ to „can we interact with the physical world.“ Companies like Physical Intelligence, Starcloud, Echodyne, and the wave of autonomous systems startups are building in a domain where the moats look nothing like software AI — manufacturing, hardware design, regulatory approval, and supply chains. You can’t GitHub clone a robot factory. There’s a version of this story where the pure-software AI crowd gets caught off guard by how much value migrates toward the intersection of intelligence and atoms, precisely because that’s where the unglamorous barriers to entry still exist.



  • Security, safety, and governance. This category is still in its infancy today, which is the point. When every company goes from a handful of AI tools to dozens of agents, the pain shifts from access to control — governing agent behavior, auditing outputs, managing security and compliance. Think about what happened in cloud: nobody cared about cloud security when companies had three workloads. When cloud became ubiquitous, Palo Alto Networks and CrowdStrike built massive businesses around securing it. AI governance — companies building the equivalent of model-level audit trails, agent access controls, output monitoring — is on that same curve, just earlier.



  • Categories that don’t even exist yet i.e the big question mark. Automobiles and commercial aviation came 30+ years after oil was struck. Radio and TV arrived decades after Edison’s first power station opened. The internet took more than a decade to explode. The truth is, the categories that will generate the most value in the age of AI haven’t even been discovered yet. Which makes this whole era even more exciting!


What Are You Building For?

 

Today, the most scarce resource in the AI supercycle is compute – spanning GPUs, memory, bandwidth, data centers, and energy. That single bottleneck is driving billions in record profits and soaring stock prices for public companies building around this layer, as well as the many private inference and GPU optimization startups collectively growing like a weed. But if prior cycles teach us any lessons, its that patterns in early innovation waves are temporary. Oil refining was scarce until it wasn’t. Bandwidth was scarce until the telecom bust accidentally created the abundance. Cloud infrastructure was scarce until AWS turned it into a utility.

When the scarcity breaks, the effects are real. Pricing power shifts. Margins compress at the resource layer. Value moves up the stack. None of this means the compute companies disappear — Exxon is still enormous and the hyperscalers print money. But the outsized returns tend to come from the companies that were building for what abundance makes possible, not from the ones optimizing around what scarcity made painful.

The question for founders and investors is a simple one: are you building for the scarcity era, or the abundance era? Because the flip is coming.

Source: https://aspiringforintelligence.substack.com/p/from-scarcity-to-abundance

Das Drei-Klassen-Netz: Technische Analyse der Mobilfunk-Capabilities im Hutchison-Universum

1 Antwort

Ein Beitrag von dieIdee.eu – Die InnovationsAgentur | Wien, März 2026

Zusammenfassung

Hutchison Drei Austria (H3A) betreibt eines der drei österreichischen Mobilfunknetze und vermarktet dieses über eigene Marken sowie externe MVNOs (Mobile Virtual Network Operators). Wer „im Drei‑Netz“ surft und telefoniert, tut dies je nach Marke jedoch unter fundamental unterschiedlichen technischen Bedingungen – hinsichtlich Frequenzzugang, 5G‑Modus, Netzpriorisierung und Zugang zum LTE‑Regional‑Roaming mit Magenta. Diese Unterschiede sind in den öffentlich zugänglichen Vertragsunterlagen für Endkunden nicht transparent ausgewiesen.

Der österreichische MVNO‑Markt: Kontext

Der österreichische Mobilfunkmarkt besteht aus drei Netzbetreibern (A1, Magenta, Drei) und einer Vielzahl virtueller Anbieter. Der MVNO‑Marktanteil liegt laut RTR bei rund 18% der SIM‑Karten (ohne M2M), HoT und spusu dominieren dieses Segment. 2023 vereinten 34 MVNOs und Submarken rund 15% Marktanteil und knapp 250 Mio. Euro Umsatz auf sich – bei einem Gesamtmarktvolumen von knapp 4 Mrd. Euro.​

Eine Studie von Arthur D. Little (ADL) zeigt: Auf einen erfolgreichen MVNO kommen fünf, die scheitern – und 50, die es gar nicht erst auf den Markt schaffen. Während MNOs in Österreich EBITDA‑Margen knapp unter 30% erzielen, gilt im MVNO‑Geschäft eine Marge von 15–20% bereits als Erfolg. Rund 60% der Gesamtkosten eines MVNO entfallen auf Access‑Kosten, also die Nutzung der fremden Netzinfrastruktur.​

Light‑MVNO vs. Full‑MVNO

Die Branche unterscheidet zwei grundlegende Modelle:​

  • Light‑MVNO: Kauft definierte Wholesale‑Pakete beim MNO, betreibt kein eigenes Core‑Netz, hat nur begrenzten Einfluss auf Preisgestaltung, technische Parameter und Gebührenstruktur, fokussiert auf Marke, Vertrieb, Kundenbeziehung.
  • Full‑MVNO: Betreibt ein eigenes Kernnetz (Core), eigene OSS/BSS, kauft nur Funkzugang beim MNO, kann Produkte, Tariflogik und technische Features flexibler gestalten.

In Österreich war Red Bull Mobile einer der ersten Full‑MVNOs; aktuell ist spusu der wichtigste Full‑MVNO im Drei‑Netz.​

Hutchison Drei Austria: Marken, MVNOs und Markenkooperationen

Unter dem Dach von Hutchison Drei Austria (FN 140132b) existiert ein vielschichtiges Marken‑ und Partnergefüge.

Eigenmarken und Markenkooperationen von Drei

  • Drei – Hauptmarke mit Premium‑ und Volumentarifen.​
  • up3 – SIM‑only‑Submarke, seit Juli 2023; Nutzungsklasse „Mobil: D“ laut Entgeltbestimmungen.​
  • Hörbi – Diskontmarke, gestartet am 4. März 2026; betrieben durch Hutchison Drei Austria, Nutzungsklasse Mobil D, siehe Entgeltbestimmungen.
  • educom – Studenten‑ und Bildungstarife.
  • eety – Ethno‑Discounter.
  • Joymobile – Nischenmarke.
  • kabelplus MOBILE – Eigenständiger Anbieter aus der EVN‑Gruppe, technisch als MVNO im Drei‑Netz; nutzt Drei‑Sendemasten (2G, LTE, 5G) und eigene Kabelnetze/FTTH für Festnetz und Internet.

Lidl Connect ist eine Marke in Zusammenarbeit zwischen Lidl Österreich und Drei: Lidl verantwortet Vertrieb und Vermarktung, Drei liefert Netz, Infrastruktur, Service und Mobilfunk‑Know‑how. Rechtlich agiert Lidl Connect damit als Markenkooperation/Eigenmarke auf Basis des Drei‑Netzes und nicht als eigenständiger Full‑MVNO im technischen Sinn.

MVNO‑Partner im Drei‑Netz

  • spusu – Full‑MVNO mit eigenem Core‑Netz und eigenen 5G‑Frequenzen (30 MHz im 3,5‑GHz‑Band in Niederösterreich).
  • Weitere kleinere Reseller.

MVNOs und Markenkooperationen in anderen Netzen

  • A1‑Netz: bob, yesss!, XOXO, Georg, Red Bull Mobile, Krone mobile, Lycamobile u.a.
    • Red Bull Mobile ist dabei eine Marke in Kooperation zwischen A1 und Red Bull („A1, bob, Red Bull MOBILE und yesss! stehen für höchste Qualität …“).
  • Magenta‑Netz: HoT (ca. 1,4 Mio. Kunden, 9,6% Marktanteil), S‑Budget Mobile, LIWEST Mobil, Raiffeisen mobil u.a.​

ADL‑Partner Christoph Uferer beschreibt die Rolle eines MVNO prägnant: Er sei „Manager der Marge zwischen dem Marktpreis und den Wholesale‑Kosten, die der Netzbetreiber verrechnet“.

Frequenzausstattung: Low‑Band‑Defizit bei Drei als strukturelle Achilles-Ferse

Drei verfügt über folgende Spektren:

FrequenzbandBandbreite (Brutto)TechnologieNutzung / Relevanz
700 MHz FDD (Band 28 / n28)2×10 MHzLTE, 5G NRFlächenversorgung, 5G SA am Land
900 MHz FDD (Band 8)2×5 MHz (davon effektiv ~3 MHz LTE)LTE, GSM, IoTIndoor‑Fallback, geringe Kapazität
1500 MHz SDL (n75)30 MHznur DownlinkKapazitätserweiterung in Kombination mit Low‑Band
1800 MHz FDD (Band 3)2×20 MHzLTE, GSMurbane Hauptversorgung
2100 MHz FDD (Band 1 / n1)2×20 MHzLTE, 5G NRKapazität in Städten
2600 MHz FDD (Band 7)2×25 MHz (davon derzeit 15 MHz aktiv genutzt)LTE, 5G NRKapazitätsband, Teilnutzung ​
2600 MHz TDD25 MHzLTE, 5G NRKapazitätsband
3400 MHz TDD (n78)100 MHz5G NRHauptband für 5G+ (SA) und 5G NSA ​

Strukturell kritisch ist, dass Drei kein Spektrum im 800‑MHz‑Band (Band 20) besitzt. Dieses Band ist in Österreich das Rückgrat der LTE‑Flächenversorgung und Indoor‑Abdeckung in ländlichen Regionen. Drei kann B20 nur über Regional‑Roaming mit Magenta nutzen – sofern der jeweilige Tarif bzw. die Marke für dieses Roaming freigeschaltet ist.

Dazu kommt: Das eigene 900‑MHz‑Band (B8) steht nur mit sehr geringer LTE‑Bandbreite zur Verfügung. In der Praxis sehen Kunden häufig zwar ein B8‑Signal, die Kapazität reicht aber kaum für stabile Datendienste. Nutzerberichte im LTE‑Forum deuten zudem darauf hin, dass Hörbi‑Tarife in einzelnen Regionen selbst auf das Drei‑eigene B8 faktisch keinen Zugriff haben, obwohl andere Drei‑Marken dort noch B8 nutzen.​

5G bei Drei: 5G, 5G+ und SA/NSA

Drei unterscheidet marketingseitig nicht zwischen „5G“ und „5G+“ im Sinne von SA/NSA, sondern nutzt „5G“ als Oberbegriff für 5G‑Dienste insgesamt.

  • „5G“ steht bei Drei für 5G NR generell, sowohl im Non‑Standalone‑Modus (NSA, mit LTE‑Anker) als auch im Standalone‑Modus (SA, reines 5G‑Core).
  • „5G+“ bezeichnet die 5G‑Standalone‑Produkte mit Bandbreitengarantie, technisch auf Basis des 3,5‑GHz‑Bandes n78.

Drei hat 5G‑Standalone im September 2022 zunächst für Internet‑Tarife („5G+ Fix“/„Premium Internet Tarife mit 5G+“) eingeführt. Handy‑Vertragstarife wurden erst ab März 2024 sukzessive für 5G SA freigeschaltet; davor stand SA nur für ausgewählte Daten‑Tarife und FWA‑Produkte zur Verfügung.

Im ländlichen Raum setzt Drei für 5G‑Flächenversorgung primär auf n28 (700 MHz), im urbanen Bereich auf n78 (3,5 GHz) und n1 (2100 MHz). Das 1500‑MHz‑SDL‑Band n75 spielt als Downlink‑Ergänzung eine Rolle, aber nur in Aggregation mit einem Ankerband.

Leistungsprofile und Nutzungsklassen

Seit Februar 2024 arbeitet Drei mit einem System statischer Leistungsprofile (A–D) und Nutzungsklassen (mobil/stationär), das in den Entgeltbestimmungen sowie in den EB von Wholesale‑Partnern wie kabelplus MOBILE dokumentiert ist.

LeistungsprofilPriorität im Netz (mobil/stationär)Beispiel‑Tarife
A600 / 300Drei Unlimited L, XL (Premium)
B300 / 150ausgewählte Premium‑Vertragstarife
C100 / 50Drei Unlimited M, Basic S
D60 / 30up3, Hörbi, MyLife SIM Start, Discountprodukte

Bei ausgelastetem Sektor erhält ein Kunde mit Profil A die sechs‑ bis zehnfache Bandbreite eines Kunden mit Profil C bzw. D. Mobile Tarife werden gegenüber stationären Tarifen zusätzlich um den Faktor 2 bevorzugt.

Die Entgeltbestimmungen von up3 weisen explizit „Nutzungsklasse: Mobil, Leistungsprofil: D“ aus. Für Hörbi ist diese Information zunächst nicht in den Vertragszusammenfassungen sichtbar, findet sich aber in den Entgeltbestimmungen: Hörbi‑Tarife sind ebenfalls der Nutzungsklasse Mobil und dem Leistungsprofil D zugeordnet.

TKK‑Bescheid C 1/23: Active Sharing, Regional Roaming und MVNO‑Auflage

Mit Bescheid C 1/23 genehmigte die Telekom‑Control‑Kommission (TKK) eine Kooperation zwischen Drei und Magenta zur gemeinsamen Nutzung aktiver Netzkomponenten („Active Sharing 1“) und zur Einführung eines strukturierten Regional‑Roaming‑Modells im 4G‑ und 5G‑Bereich.

Kernstück ist die MVNO‑Auflage, wonach Drei und Magenta die im Rahmen der Kooperation erzielten Verbesserungen bei Versorgung, Qualität und Performance „allen Endverbrauchern“ und insbesondere auch allen bestehenden MVNO‑Partnern über deren bestehende Vorleistungsverträge zugänglich machen müssen. Dies umfasst insbesondere:

  • Zugang zu Sprach‑, SMS‑ und Datendiensten in den durch Active Sharing/Regional Roaming zusätzlich versorgten 4G‑ und 5G‑Gebieten.
  • Zugang zu VoLTE‑Diensten im Netz des jeweiligen Host‑Betreibers.​

Die TKK reagierte damit auf wettbewerbsrechtliche Bedenken der Bundeswettbewerbsbehörde und des Bundeskartellanwalts, die gefordert hatten, MVNOs nicht von den Kooperationsvorteilen auszuschließen.

Wichtig: Die Auflage erwähnt ausdrücklich „MVNO‑Partner“. Hörbi ist keine MVNO‑Partnerin, sondern eine Eigenmarke von Hutchison Drei Austria (identische Firmenbuchnummer, identischer Rechtsträger). Ob drei‑interne Submarken (up3, Hörbi, educom, eety) rechtlich denselben Zugang wie externe MVNOs erhalten müssen, ist damit nicht klar geregelt.

Vergleichsmatrix der technischen Capabilities

Die folgende Tabelle fasst die wesentlichen technischen Parameter ausgewählter Marken im Drei‑Netz zusammen (Stand März 2026, auf Basis öffentlich zugänglicher Unterlagen und Nutzerberichte):

AnbieterTyp5G SA5G NSARegional Roaming Magenta (B3/B20)Nutzungsklasse / ProfilFrequenzen (praktisch)
AnbieterTyp5G SA5G NSARegional Roaming Magenta (B3/B20)Nutzungsklasse / ProfilFrequenzen (praktisch)
Drei Unlimited L/XLEigenmarkeJa (5G+)JaJamobil AB3, B1, B7(15 MHz), n28, n78
Drei Unlimited M / Basic SEigenmarkeJaJaJamobil CB3, B1, B7(15 MHz), n28, n78
up3Drei‑SubmarkeJa (ab 3/2024)JaJamobil DB3, B1, B7(15 MHz), n28, n78
Lidl ConnectMarkenkooperation Drei/LidlNein (kein 5G+)JaJa (MVNO‑Auflage)nicht publiziert, faktisch mobil C/DB3, B1, B7(15 MHz), B20 via Magenta, n78
spusu 5GFull‑MVNONein (über Drei‑Netz)JaJa (MVNO‑Auflage)nicht publiziertB3, B1, B7(15 MHz), B20 via Magenta, n78; teilweise eigene 3,5 GHz‑Zellen
kabelplus MOBILEMVNONein (5G via Drei‑NSA)JaJa (laut EB + MVNO‑Auflage)laut EB: stationär B/CB3, B1, B7(15 MHz), B20 via Magenta, n78
Hörbi 200 5G‑PowerDrei‑DiskontmarkeNein (nur 5G NSA)JaNein (kein Magenta‑Roaming)mobil DB3, B1, B7(15 MHz); B8 teils faktisch nicht nutzbar; kein B20
Hörbi 60 (LTE)Drei‑DiskontmarkeNeinNein (nur LTE)Neinmobil DB3, B1, B7(15 MHz); B8 teils faktisch nicht nutzbar; kein B20

Bei Hörbi ist besonders kritisch: Nutzer im LTE‑Forum dokumentieren Fälle, in denen selbst das Drei‑eigene B8 (900 MHz) für Hörbi‑Karten nicht nutzbar ist, während andere Drei‑Marken noch via B8 versorgt sind. In Kombination mit dem fehlenden Magenta‑Roaming führt dies dazu, dass Hörbi‑Kunden in ländlichen Regionen deutlich häufiger „Kein Netz“ sehen, obwohl andere Drei‑Produkte am selben Standort noch funktionale Versorgung haben.

Das Drei-Klassen-System

Klasse 1 – Vollzugang (Drei-Vertragstarife ab Basic S): 5G SA + NSA, volles Magenta Regional Roaming (Band 3, Band 20), Leistungsprofile A–D (je nach Tarif), alle Drei-eigenen Frequenzen einschließlich n28 und n75 im SA-Modus.

Klasse 2 – Eingeschränkter Zugang (MVNOs: Lidl Connect, spusu, kabelplus): 5G NSA (kein SA), Magenta Regional Roaming aktiv (über TKK-Bescheid C 1/23), nicht publiziertes Leistungsprofil, kein Zugang zu n28/n75 im SA-Modus. Innerhalb dieser Klasse hat spusu als Full-MVNO mit eigenem Core-Netz und eigenen 5G-Frequenzen (30 MHz im 3,5-GHz-Band in NÖ) potenziell mehr Handlungsspielraum als Light-MVNOs.

Klasse 3 – Minimalzugang (Hörbi): 5G NSA (beim 5G-Tarif) oder nur LTE (Hörbi 60), kein Magenta Regional Roaming, nur Drei-eigene Sender (B8/3 MHz, B3/15 MHz, B1/20 MHz, B7/20 MHz), kein Band 20, kein Band 28/n28 im SA-Modus, Nutzungsklasse mobil D.​

Die Konsequenz für Endkunden

Das Szenario: Drei Personen mit „5G-Verträgen im Drei-Netz“ am selben Ort mit demselben Endgerät:

  1. Drei-Direktkunde (up3): 5G SA mit n28 + n75, stabile Versorgung
  2. Lidl-Connect-Kunde (5G NSA): 4G über Magenta Regional Roaming (Band 20), brauchbarer Empfang
  3. Hörbi-Kunde (5G NSA, kein Magenta): Kein Empfang, EDGE, oder ein Balken ohne Konnektivität

ADL-Partner Christoph Uferer beschreibt die Rolle eines MVNO als „Manager der Marge zwischen dem Marktpreis und den Wholesale-Kosten, die der Netzbetreiber verrechnet“. Das gilt für echte MVNOs. Bei Hörbi als Eigenmarke des Netzbetreibers stellt sich die Frage anders: Hier entscheidet der Netzbetreiber selbst, welchen Teil seiner Infrastruktur er seiner Diskontmarke zur Verfügung stellt – und wie transparent er das kommuniziert.​

Regulatorische Perspektive

Die RTR stellte in ihrem Newsletter 02/2025 fest:​

  • Der MVNO-Anteil bei 5G-SIM-Karten liegt „deutlich unter 5%“, während der Gesamtmarktanteil rund 18% beträgt
  • „Bei Technologieübergängen – etwa zu 5G, oder auch zu VoLTE – wurde der Zugang zur Netzinfrastruktur für MVNOs oftmals verzögert oder nur mit höheren Vorleistungsentgelten gewährt“
  • „Bei VoLTE ermöglichte bei einzelnen Anbietern erst eine Auflage in einem regulatorischen Verfahren den Zugang“

Die eSIM-Technologie wird den MVNO-Markt weiter dynamisieren. Laut ADL stehen 30 bis 40% der österreichischen Smartphones bereits mit eSIM-Funktion zur Verfügung. Der digitale Point-of-Sale wird physische Filialen ersetzen – ADL-Partner Uferer spricht von einer „totalen Revolution“. Für 2026 ist zudem der MVNO-Launch der Österreichischen Post im A1-Netz angekündigt.​

In diesem dynamischen Marktumfeld wäre eine regulatorische Klarstellung überfällig: Müssen die technischen Capabilities eines Tarifs – 5G SA/NSA, Regional-Roaming-Zugang, Leistungsprofil – für Endkunden transparent ausgewiesen werden? Die aktuelle Informationslage ist unzureichend. Weder in Hörbi-Vertragszusammenfassungen noch in den Entgeltbestimmungen von Lidl Connect finden sich Angaben dazu.

Quellen: RTR Newsletter 02/2025; TKK-Bescheid C 1/23 (Kooperationsgenehmigung Drei/Magenta); RTR-Presseaussendung 12.1.2024; Entgeltbestimmungen Drei (Special Tarife, April 2025; up3, Mai 2025; Unlimited Tarife, Januar 2025); Vertragszusammenfassungen Hörbi (Februar 2026); Entgeltbestimmungen Lidl Connect (Februar 2026); Leistungsbeschreibung spusu 5G (Jänner 2024); OTS-Presseaussendung Hörbi (4.3.2026); Arthur D. Little, „Riding the MVNO Wave: 10 Keys to Success“; report.at, „MVNO – Ein schwieriger Markt“ (2025); tarife.at; teltarif.de; lteforum.at.

Traditional AdTech is Dead. Long Live AdTech For AI

Kommentar verfassen

The rebirth of advertisements and AdTech in the age of AI

AdTech used to matter. In the early 2000s, it was one of the hottest areas in tech – hundreds of startups, billions in VC funding, genuine innovation in targeting and formats. Then Google bought DoubleClick in 2007 for $3.1 billion, Facebook launched its ad platform, and the game was over. The duopoly that emerged didn’t just dominate—at their peak, Google and Meta controlled nearly 80% of U.S. digital ad growth. Everyone else was left fighting for scraps. For the past 15 years, „AdTech startup“ became practically an oxymoron as the industry consolidated into irrelevance.

But now, in the age of AI, we are starting to see a resurgence of advertising as a booming revenue source for companies. OpenAI announced last week that they would be testing ads in ChatGPT in a “bid to boost revenue”, and the healthcare AI startup OpenEvidence recently surpassed $100M annualized run-rate revenue (and doubled their valuation to $12B!), largely on an ad-supported revenue model. And around this, the market for AI tools in advertising optimization is growing quickly.

So why are we seeing such a sharp resurgence in a field that just a few years ago was essentially dead?

Three fundamental shifts are driving this renaissance:

First, AI platforms have created the first genuinely new advertising surface since social media. ChatGPT’s 800+ million weekly active users and Claude’s ~20 million monthly active users represent massive, engaged audiences that didn’t exist two years ago. Unlike the incremental improvements of the past decade – slightly better targeting, marginally improved attribution – these platforms represent entirely new real estate where the old duopoly rules don’t apply.

Second, intent signals are dramatically more sharply defined than anything we’ve seen before. When someone types “best CRM for startups” into Google, you get a decent intent signal. But when someone has a 20-message conversation with ChatGPT about their specific sales team structure, pain points, budget constraints, and technical requirements? That’s intent data at a resolution advertisers have only dreamed about. The conversational nature of AI interactions creates a richness of context that search queries simply can’t match.

Third, entirely new infrastructure is required—and being built at breakneck speed. The old AdTech stack was built for display ads, search results, and social feeds. None of it works for conversational AI. How do you measure attribution when there’s no “click”? How do you bid on inventory that’s generated dynamically in response to natural language? What does “viewability” even mean in a text-based conversation? This infrastructure gap is spawning entirely new categories like Generative Engine Optimization (GEO), with dozens of startups raising millions to solve problems that didn’t exist 18 months ago.

So should we be heralding the rebirth of AdTech in the age of AI?

Let’s dig in.

AI is Creating New Real Estate for Ads

The rapid growth of foundation models with “chatbot-style interfaces” has brought forward what we believe is the first new real estate for advertisements since the emergence of social networks. Google and Meta were able to establish dominance in ad models by aggregating eyeballs; Google in search and Meta in social. (And to a lesser extent other social media platforms like Snap, Pinterest, etc.). As an advertiser, why would I place my ads anywhere other than where consumers are aggregating to get most bang for my buck?

Now in the Age of AI, consumers are no longer flocking to the traditional platforms but increasingly to places like ChatGPT (>800M WAUs) and Claude (~20M MAUs). And these consumers are not just making simple search queries but having full-blown conversations on every topic under the sun. This is an advertiser’s dream: a wide-scale canvas with rich, user-generated intent. And advertisers are no longer limited to traditional search displays with sponsored results but can embed more natural advertisements within AI-generated responses. While the consumer may not love this (more on that below), it certainly makes sense for the advertisers.

Beyond consumer-facing platforms, AI development tools are creating a quieter but equally significant advertising opportunity. When developers use tools like Lovable, Replit, or Cursor to build applications, these platforms make dozens of architectural decisions on their behalf—which database to use, where to host, which payment processor to integrate.

Each of these decisions represents potential advertising inventory. Supabase could sponsor recommendations in database selection flows. Vercel could appear as a ‘suggested deployment option’ when a developer’s app is ready to ship. Stripe could surface contextual offers when payment processing code is being written.

The key difference from traditional developer advertising (think Stack Overflow banner ads) is that these aren’t interruptions—they’re recommendations at the exact moment of intent. A developer isn’t being shown a database ad while reading about React hooks; they’re being offered database options precisely when their AI agent is about to scaffold database code. The conversion potential is orders of magnitude higher.

Vertical AI Is Creating Specialized Inventory

It’s not just the large model providers themselves that are benefiting from ads.

The rise of vertical AI providers is creating a new, specialized inventory for high-intent, high-value ads. Verticals like healthcare, legal, finance, real estate, and other professional services are becoming the new adtech frontier, offering advertisers direct access to high-value audiences outside the Google-Meta duopoly for the first time in over a decade.

One great example here is OpenEvidence, which has quickly grown into the leading “AI-powered medical search engine” for clinicians. The company recently said that 40% of physicians across the US across 10K hospitals and medical centers now use OpenEvidence on a daily basis. What else is interesting and unique is its business model: OpenEvidence is free to use for verified medical professionals, and generates revenue largely through advertising.

Per a great business breakdown from Contrary Research:

Given that pharmaceutical companies spent approximately $20 billion annually on marketing to healthcare professionals in the US as of 2019, capturing a portion of this market through digital channels could generate substantial revenue for the company. OpenEvidence’s advertising focus on contextual advertising and sponsored content while maintaining trust. For example, if a doctor submits a query about diabetes treatments, a sponsored summary from a pharmaceutical manufacturer may appear, or a banner for relevant clinical webinars could be displayed.

This advertising model has allowed OpenEvidence to reach >$100M annualized run-rate revenue in just a few short years.

We believe that other vertical AI tools will also embed this type of model, giving away the product to end users for free while generating revenue from charging advertisers. In vertical AI, the intent signals are clearer than ever—and unlike the generic search box, users are getting AI agents that actually solve their specific problems, creating a sustainable value exchange that justifies the ad-supported model.

Measurement Primitives are Changing and New Infrastructure is Emerging

Attribution in AI-native experiences is fundamentally different. The old AdTech stack was built for discrete surfaces where ads could be served, clicked, and tracked, but in conversational and agentic interfaces, there’s often no obvious “ad slot” and no click at all. Instead, influence is embedded inside multi-turn workflows: what the model recommends, what the user accepts, and what gets generated.

In next-gen AI apps, advertising is moving into the flow of work. When a developer scaffolds an app in Cursor, Lovable, or Vercel, the inventory isn’t a banner but it’s the moment an agent suggests a database, auth provider, or cloud service. In vertical AI tools, the same pattern holds: the “ad” looks like a contextual recommendation for a pharmaceutical brand, clinical resource, or specialized service. These touchpoints are integrated into the utility itself.

This shift is spawning an entirely new measurement rail. If clicks disappear, the new primitives become telemetry and adoption: logging multi-turn conversations, mapping model outputs to downstream actions, and tracking “acceptance events” like tab-to-insert, install, purchase, or integration. And because influence in a conversation is cumulative, the real challenge isn’t just attribution but it’s incrementality: did the recommendation actually change what the user would have done otherwise?

As a result of this shift, we are starting to see new ad networks emerge to serve these “in-flow” moments.

  • On the measurement side, companies like Profound and Bluefish are building the GEO observability layer, tracking share-of-response, competitive displacement, and brand presence across models.
  • On the distribution side, a new generation of AI-native ad platforms is forming across multiple surfaces: platforms like ZeroClick, OpenAds, and Nex.ad are beginning to monetize dynamic, contextually relevant recommendations inside or alongside AI conversations, while publisher-centric AI engagement platforms like Linotype.ai help site owners retain users and surface native monetization opportunities.

But unlike the old web, the “auction” can’t just pick the highest bidder. It has to operate inside generation loops, ranking units based on contextual relevance, quality, and bid while navigating trust and policy constraints in sensitive domains like healthcare and legal. Pricing models may shift as well, away from CPM/CPC and toward outcomes like cost-per-accept, cost-per-embed, or cost-per-adoption.

The biggest wildcard is walled gardens. If OpenAI, Anthropic, Google, and vertical copilots control the interface, they may also control the inventory and measurement rails, turning AI advertising into a handful of closed ecosystems rather than an open programmatic market. Time will tell!

Nexad.ai Reinventing Ads for the AI Era

Conclusion / Challenges

There’s one clear challenge in all of this: people generally dislike advertisements. A recent report found that 81% of young people hate ads, and 60% find them intrusive. And who’s to blame them? Most people find ads annoying and not beneficial to them, and now nearly half the internet uses an ad blocker.

Another key challenge is whether people will feel that the answers they are served by the LLMs are influenced by the advertisements that appear. If I ask Claude for the best recommendations for hotels in Switzerland, will I know it showing me what the model says is “best”, or which hotel is spending the most on advertising for this query result?

But here’s the interesting part: in the same study referenced above, only 28% of respondents wanted fewer ads. Which suggests that its not the brands or products being peddled they dislike, but how the ads are actually served.

This could actually be a boon to the new platforms like OpenAI and Anthropic, as well as the emerging AI Adtech tools. By finding creative, non-intrusive, intent-based, transparent, and beneficial ways to reach consumers, a new form of advertising could actually flourish.

So we’re left with the thought…

“Traditional AdTech is Dead…Long Live AdTech For AI“.

Source: https://aspiringforintelligence.substack.com/p/traditional-adtech-is-dead-long-live

The Death of Software 2.0

Kommentar verfassen

The age of PDF is over. The time of markdown has begun. Why Memory Hierarchies are the best analogy for how software must change. And why Software it’s unlikely to command the most value.

When I last wrote about software, I received significant pushback. Today, I believe that Claude Code is confirming the original case I had all along. Software is going to become an output of hardware and an extension of current hardware designs. With this in mind, I want to write today about how I see software changing from here.

But let’s start with one core conviction. Claude Code is the glimpse of the future. Assuming it improves, has harnesses, and can continue to scale large context windows and only become marginally more intelligent, I believe this is enough to really take us to the next state of AI. I cannot stress enough that Claude Code is the ChatGPT moment repeated. You must try it to understand.

One day, the successor to Claude Code will make a superhuman interface available to everyone. And if Tokens were TCP/IP, Claude Code is the first genuine website built in the age of AI. And this is going to hurt a large part of the software industry.

Software (Especially Seat Based) is in for a Much Rougher Ride

The environment may be rough at OpenAI, but at a traditional SaaS company, there is likely no greater whiplash than SaaS is eating the world in 2012 to Saas is screwed today. The stocks reflect it; multiple compression in the companies has been painful and will persist.

Source: EODHD

This is structural. I believe it’s time to rethink software’s value proposition, and I have what I consider the best analogy for what the future holds. Afterwards, we will digest what Software will look like as an extension of computing, because I believe that Claude Code resembles the memory hierarchy in computing. Let’s explain.

The New Model of Software

Claude Code (and subsequent innovations) clearly will change a lot about software, but the typical (and right) pushback is that you cannot use “non-deterministic software” for defined business practices. However, there is a persistent design pattern in hardware that addresses this difference: the memory hierarchy. No one can rely on anything in a computer’s non-persistent memory, yet it is one of the most valuable components of the entire stack.

For those unfamiliar with computer science, there is a memory hierarchy that trades capacity and persistence for speed, and the system works because there are handoffs between levels. In the traditional stack, SRAM sits at the top; overflow is to DRAM, which is non-persistent (if you turn it off, it goes away), and then to NAND, which is persistent (if you turn it off, it persists).

I don’t think it’s worth matching the hierarchy too closely, but I believe that Claude Code and Agent Next will be the non-persistent memory stack in the compute stack. Claude Code is DRAM.

Memory Hierarchy
Source: https://computerscience.chemeketa.edu/cs160Reader/ComputerArchitecture/MemoryHeirarchy.html

I believe that AI and software will be an extension of this, and we can already identify which layers correspond to which. The “CPU” in the hierarchy comprises raw information, and the fast memory in the hierarchy corresponds to the context window. This level of context is very fast information, not persistent, and gets cleared systematically. The output of work performed in non-persistent memory is passed to the NAND, which is stored for the long term.

Now that the code is merely an output of hardware, I believe this analogy applies.

AI Agents and their context windows are going to be the new “fast memory”, and I believe that infrastructure software is going to look a lot closer to persistent memory. It will have high value, structured output, and will be accessed and transformed at a much slower rate. I believe the way to think of software, and the “software of the future,” looks a lot more like NAND, and that is persistent, accurate, and information that needs to be stored. In software parlance, it will be the “single source of truth” that AI agents will interact with and manipulate information from.

If you’re so visually inclined, here’s a diagram of a Claude code context window being compacted over and over. Another way to think about this is that it is an identity for a compute cycle; once the task is finished, it is transferred to slower memory and continues.

Source: Weka

Each time an AI agent’s computation cycle occurs, this is a scratchpad. Each context window is a clock cycle: cached state accumulates until the cache is flushed, after which information is processed. Afterward, the entire context is discarded, leaving only the output. Computation is ephemeral, and information processing by a higher tier of computation largely abstracts away most of human reasoning.

Importantly, I think there is a world in which software doesn’t go away, but its role must change. In this analogy, data, state, and APIs will be persistent storage, akin to NAND, whereas human-oriented consumption software will likely become obsolete. All horizontal software companies oriented at human-based consumption are obsolete. The entire model will be focused on fast information processors (AI Agents), using tokens to transform them and depositing the answers back into memory. Software itself must change to support this core mechanism, as the compute engine at the top of the hierarchy is primarily nonhuman, namely an AI agent.

I believe that next-generation software companies must completely shift their business models to prepare for an AI-driven future of consumption; otherwise, they will be left behind.

Glimpses of the Future

So what does this future look like? I believe that all software must leave information work as soon as possible. I believe that the future role of software will not have much “information processing”, i.e., analysis. Claude Code or Agent-Next will be doing the information synthesis, the GUI, and the workflow. That will be ephemeral and generated for the use at hand. Anyone should be able to access the information they want in the format they want and reference the underlying data.

What I’m trying to say is that the traditional differentiation metrics will change. Faster workflows, better UIs, and smoother integrations will all become worthless, while persistent information, a la an API, will become extremely valuable. Software and infrastructure software will become the “NAND” portion of the memory hierarchy.

And since I’m going to be heavily relying on the history of memory, the last time a new competitive technology came out, it was an extinction event for the magnetic cores that DRAM replaced, and I think this is probably going to be the case for UI companies or companies like Tableau or other visualization software. Zapier / Make as connectors, UiPath, or RPA companies, etc. These are all facing an extinction-level event.

Other companies that I think could be significantly affected include Notion and Airtable. Monday, Asana, and Smartsheet are merely UIs for tasks; why should they exist? Figma could be significantly disrupted if UIs, as a concept humans create for other humans, were to disappear.

Companies that are interesting are “sources of truth,” but many of them need to change. An example might even be Salesforce, a SaaS company. I don’t think the UI is that great, and most of the custom projects are just hardening workflows in the CRM. For Salesforce to make the leap, it needs to focus its product on being consumed by an AI agent, with manipulation and maintenance, while being the best possible NAND in this stack. The problem is that Salesforce will want to try to go up the stack, and by doing so, maybe miss the shift completely.

Most SaaS companies today need to shift their business models to more closely resemble API-based models to align with the memory hierarchy of the future of software. Data’s safekeeping and longer-term storage are largely the role of software companies now, and they must learn to look much more like infrastructure software to be consumed by AI Agents. I believe that is what’s next.

This raises the question: what does this look like for the industry as a whole in the near future? I believe the next 3-5 years will be a catastrophic sea change.

Source: https://www.fabricatedknowledge.com/p/the-death-of-software-20-a-better

Claude Code: If you’re still typing instructions into Claude Code like you’re asking ChatGPT for help, you’re missing the entire point

Kommentar verfassen

Source: https://diamantai.substack.com/p/youre-using-claude-code-wrong-and

If you’re still typing instructions into Claude Code like you’re asking ChatGPT for help, you’re missing the entire point. This isn’t another AI assistant that gives you code snippets to copy and paste. It’s a different species of tool entirely, and most developers are using maybe 20% of what it can actually do.

Think of it this way: you wouldn’t use a smartphone just to make phone calls, right? Yet that’s exactly what most people do with Claude Code. They treat it like a glorified autocomplete engine when it’s actually a complete development partner that lives in your terminal, understands your entire codebase, and can handle everything from architecture decisions to writing documentation.

The gap between casual users and power users isn’t about technical knowledge. It’s about understanding the workflow, knowing when to intervene, and setting up your environment so Claude delivers production-quality results consistently. This guide will show you how to cross that gap.

Your Development Partner Lives in the Terminal

Picture working with a senior developer who never gets tired, can read thousands of files in seconds, and has instant access to the entire internet. That’s Claude Code. It connects Anthropic’s AI models directly to your project through the command line. You describe what you need in plain language, and it plans solutions, writes across multiple files, runs tests, and implements features.

But here’s what makes it different from every other coding tool: it actually understands context. Not just syntax highlighting or function signatures. Real context. It reads your project structure, sees your existing patterns, runs your tools, and even fetches information from external sources when needed.

The catch is this: giving it instructions is a skill. A learnable skill, but a skill nonetheless. The difference between getting mediocre results and getting genuinely useful code comes down to how you communicate and how you structure your workflow.

The One Rule That Changes Everything

Here’s where most people go wrong immediately: they start coding right away. It’s like walking into a contractor’s office and saying “start building my house” without showing blueprints, discussing materials, or even agreeing on what kind of house you want.

The result? You’ll get a house. It might even have walls and a roof. But it probably won’t be what you imagined.

Always start in plan mode. Before giving any instructions, press shift-tab to cycle into planning mode. Tell Claude to explore your codebase first, but specifically tell it not to write anything yet. Let it read the relevant files, understand the architecture, and grasp the bigger picture.

Once it’s explored, ask for a proposal. Not the simplest solution, not the fastest solution. Ask it to think through options starting with the most straightforward approach. Then discuss that plan like you would with a colleague. Question assumptions. Refine the approach. Push back if something seems off.

Only after you’re confident it understands the task should you tell it to start coding.

This feels slower at first. Your instinct will be to just dive in and start building. Resist that instinct. Planning five minutes saves fixing broken implementations for an hour. Every single time.

Precision Beats Brevity Every Time

Vague instructions produce vague results. Say “fix the bug” and you might get a fix, or you might get a complete rewrite that breaks three other features. There’s no middle ground here.

Instead, be surgical with your instructions. Point to specific files. Reference exact functions. Mention line numbers if you have them. Compare these two approaches:

“Fix the authentication issue.”

versus

“In the login.js file in the auth folder, update the token validation function to handle expired tokens without crashing.”

The second version leaves no room for misinterpretation. It guides Claude exactly where to look and what to do.

This precision applies to style and patterns too. If you want code that matches your existing codebase, say so explicitly. Point Claude to well-written examples in your project. It can mirror patterns beautifully, but only when you show it the pattern you want.

Think of it like directing a movie. You wouldn’t tell an actor “do something emotional.” You’d say “show hesitation, then determination, with a slight smile at the end.” Same energy here.

Your Most Powerful Tool Is the Escape Key

Claude works best as a collaborative partner, not an autonomous robot. The Escape key keeps you in control.

See Claude heading down the wrong path? Hit Escape immediately. This stops it mid-process while keeping all the context intact. You can redirect without losing the work already done. It’s like tapping someone on the shoulder mid-sentence and saying “wait, different approach.”

Double-tap Escape to jump backward through your conversation history. This lets you edit a previous prompt and explore an alternative direction. You can iterate on the same problem multiple times, trying different solutions until one clicks.

If Claude makes changes you don’t like, just tell it to undo them. It can revert files instantly. Combined with regular checkpoints, this means you can experiment fearlessly. The safety net is always there.

Understanding the Different Modes

Claude Code has multiple modes, and knowing when to use each one separates beginners from experts.

Plan mode is for thinking, not doing. Use it when starting new features or untangling complex problems. It will architect solutions without touching your files. This is your strategy phase.

Code mode is for building. Once you have a solid plan, switch to code mode and let it implement. But stay alert. Watch what it’s doing and be ready to course-correct.

Auto-accept mode removes the approval step for each change. It’s fantastic for straightforward tasks but dangerous for anything complex or important. For critical work, stay manual and review everything.

Bash mode lets you run terminal commands and feed the output directly into Claude’s context. This is debugging gold. Run your tests, capture the failures, and immediately ask Claude to fix them without copying error messages around.

Each mode has its place. The trick is recognizing which situation calls for which mode.

Managing Context Before It Manages You

Claude Code’s biggest weakness is context window limits. As sessions grow longer, it starts forgetting earlier information. Power users have strategies to handle this.

Use the /compact command regularly. It clears old execution results while keeping the important conversation history. Think of it like cleaning your desk: you keep the critical documents but toss the scrap paper.

For complex projects, create a CLAUDE.md file in your project root. This becomes permanent memory. Put your project overview, architecture decisions, coding standards, and common patterns there. Claude reads it automatically and uses it as context for every task. It’s like giving every session a primer on how your project works.

For massive tasks, use a checklist file. Create a markdown document with all the steps needed to complete the task. Tell Claude to use it as a scratchpad, checking off items as it progresses. This keeps the main conversation clean while giving Claude a progress tracker.

Divide Complex Work with Subagents

When facing a genuinely complex problem, break it apart and assign pieces to different subagents. Tell Claude to spin up a subagent for the backend API while the main agent handles the frontend. Or have one subagent research documentation while another writes implementation code.

You can even mention subagents directly with the @ symbol to guarantee they activate. You can also specify which model each subagent should use. Opus 4 handles complex planning and architecture. Haiku 3.5 knocks out simple, fast tasks.

This approach tackles problems in parallel and keeps context focused. Each subagent deals with one slice of the problem without getting overwhelmed by the full complexity. It’s like having multiple specialists working on different parts of a project simultaneously.

Show, Don’t Tell

Claude Code can interpret visual information. Drag screenshots directly into your terminal. Show it UI mockups, error messages, or architecture diagrams. It will understand the visual context and use it to guide implementation.

This is especially powerful for debugging interface issues. Instead of describing what’s wrong with your layout, just show a screenshot. For replicating designs, provide the mockup and let Claude figure out the implementation details.

Visual context often communicates more than words ever could. A single screenshot can replace three paragraphs of explanation. Use this liberally.

Automate Everything, Then Automate the Automation

Claude Code excels at repetitive tasks. But power users go further: they automate the automation itself.

Set up custom slash commands for tasks you repeat constantly. Create a command that loads your project context, runs your test suite, and generates documentation in one go.

Use hooks to trigger actions automatically. Run tests after every code change. Lint before commits. Update documentation when finishing features. These small automations compound into massive time savings.

For data processing pipelines, integrate Claude directly into your workflow. Pipe data in, let it transform or analyze the data, and pipe the output to the next step. This turns Claude into a powerful processing node in your toolchain.

Extended Thinking for Complex Problems

For genuinely difficult problems, use extended thinking commands like /think or /ultrathink. These increase Claude’s reasoning budget, giving it more time to work through complicated challenges.

Yes, it takes longer. But the quality difference is dramatic for debugging, architecture planning, and design decisions. It’s the difference between asking for a quick answer versus asking someone to really think through a problem thoroughly.

The ultrathink command is particularly powerful. It provides the maximum thinking budget, perfect for architectural decisions or bugs that have stumped you for hours. Use it sparingly, but when you need it, you really need it.

Git Workflows That Keep You Safe

Never work directly on your main branch with Claude Code. Always create a feature branch first. This gives you a safe sandbox to experiment in.

Even better, use Git worktrees. This lets you maintain multiple working directories for different branches, so you can have Claude working on several features in parallel without interference.

When Claude finishes a task, have it commit changes with a clear message explaining what was done. Then review the commit diff carefully before merging. This workflow gives you the safety of version control while letting Claude work autonomously.

Embed Your Standards in Documentation

Instead of reminding Claude about coding standards in every conversation, embed them directly in documentation files. Create a QUALITY.md file with your coding standards, testing requirements, and review checklist.

Claude will read this automatically and follow your standards without being told. It becomes part of the project context, like a senior developer who knows the house rules and follows them instinctively.

For teams, this ensures consistency across all Claude Code sessions. Everyone gets the same quality bar, regardless of who’s running the tool.

The MCP Revolution

Model Context Protocol servers extend Claude Code’s capabilities dramatically. Connect it to your Slack, Figma, Google Drive, or custom data sources. This transforms Claude from a code assistant into a genuine team member that can pull information from anywhere.

Need to check the latest design mockup? Claude fetches it from Figma. Need to understand a business requirement? It pulls it from Slack.

Set up MCP servers for your most-used tools. The initial setup takes time, but the payoff is enormous. Claude becomes infinitely more capable when it can access your actual data sources.

Debugging Strategy

Claude Code is exceptional at debugging when you give it proper information. When you hit a bug, don’t just paste the error message. Use bash mode to run your tests and feed the full output to Claude. Tell it to analyze the stack trace, read the relevant files, and propose a fix.

For intermittent bugs, run the failing code multiple times and give Claude all the outputs. It can spot patterns in failures that humans miss.

If bugs involve external services, use Claude to fetch relevant documentation or logs. It can correlate error messages with API documentation to pinpoint exactly what’s wrong.

Self-Writing Documentation

One of Claude Code’s most underrated features is documentation generation. After finishing a feature, tell Claude to update the README, API docs, and changelog. It has full context of what was just built, so it writes accurate, comprehensive documentation without requiring explanation.

This is especially powerful after refactors, where documentation typically gets forgotten. Set up a hook to automatically generate documentation after every feature merge. Your docs will stay synchronized with your code effortlessly.

Managing Token Usage in Long Sessions

Long Claude Code sessions can get expensive as context grows. Smart users manage this proactively.

Break large tasks into smaller chunks. Complete one chunk, commit it, then start a fresh session for the next chunk. This keeps context size manageable and costs reasonable.

Use prompt caching for information that doesn’t change often. Load your project overview and standards once, then reference them in subsequent sessions. This dramatically reduces token usage.

For repetitive tasks across many files, use a script to process them in batches rather than one giant session. This parallel approach is both faster and more cost-effective.

The Checklist Method for Large Migrations

For migrations, massive refactors, or fixing hundreds of lint errors, the checklist method is unbeatable.

Create a markdown file listing every task that needs completion. Tell Claude to use this as its working document, checking off items as it completes them and adding notes about any issues.

This approach does two crucial things. First, it gives Claude a clear roadmap, preventing it from getting lost in complexity. Second, it lets you track progress and see exactly what’s been done.

For truly large codebases, break the checklist into sections and tackle them in separate sessions. This keeps each session focused and productive.

Accelerating Learning and Onboarding

Claude Code is an incredible learning tool. New team members can ask it to explain the codebase, trace through execution flows, and understand architecture decisions.

Have newcomers ask Claude to map out the project structure and identify key components. Then they can ask specific questions about how things work. This accelerates onboarding from weeks to days.

For existing team members exploring unfamiliar parts of the codebase, Claude provides guided tours. Ask it to explain the authentication flow or the data pipeline, and it will trace through the code, explaining each piece clearly.

Beyond the Code

Claude Code can do much more than write software. Use it for research tasks, like reading documentation and creating summaries for future reference. It can analyze logs, process data files, and generate reports.

Need to understand a new API? Have Claude read the documentation and create a usage guide. Working with a large CSV file? Pipe it into Claude and ask for analysis.

These non-coding tasks often consume huge amounts of developer time. Claude can handle them while you focus on the creative problem-solving that actually requires human intelligence.

Avoiding Common Traps

Even experienced users make mistakes. Here are the most frequent ones and how to sidestep them.

Trusting auto-accept mode for complex tasks is dangerous. Auto-accept is convenient but risky for anything affecting core functionality. Always review changes manually for important work.

Letting sessions run too long accumulates context and makes everything slower and more expensive. Refresh your session regularly, especially after completing major milestones.

Not using version control is asking for trouble. Always use branches, and always review diffs before merging.

Being too vague leads to assumptions. Those assumptions might not match your intent. Take time to be precise.

Ignoring the plan phase might feel faster, but it leads to rework. The few minutes spent planning save hours of fixing wrong implementations.

Bosses are fighting a new battle in the RTO wars: It’s not about where you work, but when you work

Kommentar verfassen

For the last three years, the corporate world has been locked in a territorial dispute. The “Return to Office” (RTO) wars were defined by geography: the home versus the headquarters. But as 2025 unfolded, the frontline shifted. According to commercial real-estate giant JLL’s Workforce Preference Barometer 2025, the most critical conflict between employers and employees is no longer about location—it is about time.

While structured hybrid policies have become the norm, with 66% of global office workers reporting clear expectations on which days to attend, a new disconnect has emerged. Employees have largely accepted the “where,” but they are aggressively demanding autonomy over the “when.”

The report highlights a fundamental change in employee priorities. Work–life balance has overtaken salary as the leading priority for office workers globally, cited by 65% of respondents—up from 59% in 2022. This statistic underscores a profound shift in needs: Employees are looking for “management of time over place.”

While high salaries remain the top reason people switch jobs, the ability to control one’s schedule is the primary reason they stay. The report notes employees are seeking “agency over when and how they work,” and this desire for temporal autonomy is reshaping the talent market.

Although JLL didn’t dive into the phenomenon of “coffee badging,” its findings align with the practice of hybrid workers stretching the boundaries of office attendance. The phrase—meaning when a worker badges in just long enough to have the proverbial cup of coffee before commuting somewhere else to keep working remotely—vividly illustrates how the goalposts have shifted from where to when. Gartner reported 60% of employers were tracking employees as of 2022, twice as many as before the pandemic.

The ‘flexibility gap’

JLL’s data reveals a significant “flexibility gap”: 57% of employees believe flexible working hours would improve their quality of life, yet only 49% currently have access to this benefit.

The gap is particularly dangerous for employers, JLL said, arguing it believes the “psychological contract” between workers and employers is at risk. While salary and flexibility remain fundamental to retention, JLL said its survey of 8,700 workers across 31 countries reveals a deeper psychological contract: “Workers today want to be visible, valued and prepared for the future. Around one in three say they could leave for better career development or reskilling opportunities, while the same proportion is reevaluating the role of work in their lives.” JLL argued “recognition … emotional wellbeing and a clear sense of purpose” are now crucial for long-term retention.

The report warns that where this contract is broken, employees stop engaging and start seeking compensation through “increased commuting stipend and flexible hours.” The urgency for time flexibility is being driven by a crisis of exhaustion. Nearly 40% of global office workers report feeling overwhelmed, and burnout has become a “serious threat to employers’ operations.”

The link between rigid schedules and attrition is clear: Among employees considering quitting in the next 12 months, 57% report suffering from burnout. For caregivers and the “squeezed middle” of the workforce, standard hybrid policies are insufficient; 42% of caregivers require short-notice paid leave to manage their lives, yet they often feel their constraints are “poorly understood and supported at work.”

To survive this new battle, the report suggests companies must abandon “one-size-fits-all” approaches. Successful organizations are moving toward “tailored flexibility,” which emphasizes autonomy over working hours rather than just counting days at a desk. This shift even impacts the physical office building. To support a workforce that operates on asynchronous schedules, offices must adapt with “extended access hours,” smart lighting, and space-booking systems that support flexible work patterns rather than a rigid 9-to-5 routine.

Management guru Suzy Welch, however, warns it may be an uphill battle for employers to find a burnout cure. The New York University professor, who spent seven years as a management consultant at Bain & Co. before joining Harvard Business Review in 2001, later serving as editor-in-chief, told the Masters of Scale podcast in September burnout is existential and generational. The 66-year-old Welch argued burnout is linked to hope, and current generations have reason to lack this.

“We believed that if if you worked hard you were rewarded for it. And so this is the disconnect,” she said.

Expanding on the theme, she added: “Gen Z thinks, ‘Yeah, I watched what happened to my parents’ career and I watched what happened to my older sister’s career and they worked very hard and they still got laid off.’” JLL’s worldwide survey suggests this message has resonated for workers globally: They shouldn’t give up too much of their time, because it just may not be rewarded.

Source: https://fortune.com/2026/01/04/bosses-new-battle-remote-work-from-where-to-when-flexible-hybrid

Agentic Search and The Search Wars of 2026: ChatGPT’s Conversational Surge Challenges Google’s Decades-Long Hegemony

Kommentar verfassen

As of January 2, 2026, the digital landscape has reached a historic inflection point that many analysts once thought impossible. For the first time since the early 2000s, the iron grip of the traditional search engine is showing visible fractures. OpenAI’s ChatGPT Search has officially captured a staggering 17-18% of the global query market, a meteoric rise that has forced a fundamental redesign of how humans interact with the internet’s vast repository of information.

While Alphabet Inc. (NASDAQ: GOOGL) continues to lead the market with a 78-80% share, the nature of that dominance has changed. The „search war“ is no longer about who has the largest index of websites, but who can provide the most coherent, cited, and actionable answer in the shortest amount of time. This shift from „retrieval“ to „resolution“ marks the end of the „10 blue links“ era and the beginning of the age of the conversational agent.

The Technical Evolution: From Indexing to Reasoning

The architecture of ChatGPT Search in 2026 represents a radical departure from the crawler-based systems of the past. Utilizing a specialized version of the GPT-5.2 architecture, the system does not merely point users toward a destination; it synthesizes information in real-time. The core technical advancement lies in its „Citation Engine,“ which performs a multi-step verification process before presenting an answer. Unlike early generative AI models that were prone to „hallucinations,“ the current iteration of ChatGPT Search uses a retrieval-augmented generation (RAG) framework that prioritizes high-authority sources and provides clickable, inline footnotes for every claim made.

This „Resolution over Retrieval“ model has fundamentally altered user expectations. In early 2026, the technical community has lauded OpenAI’s ability to handle complex, multi-layered queries—such as „Compare the tax implications of remote work in three different EU countries for a freelance developer“—with a single, comprehensive response. Industry experts note that this differs from previous technology by moving away from keyword matching and toward semantic intent. The AI research community has specifically highlighted the model’s „Thinking“ mode, which allows the engine to pause and internally verify its reasoning path before displaying a result, significantly reducing inaccuracies.

A Market in Flux: The Duopoly of Intent

The rise of ChatGPT Search has created a strategic divide in the tech industry. While Google remains the king of transactional and navigational queries—users still turn to Google to find a local plumber or buy a specific pair of shoes—OpenAI has successfully captured the „informational“ and „creative“ segments. This has significant implications for Microsoft (NASDAQ: MSFT), which, through its deep partnership and multi-billion dollar investment in OpenAI, has seen its own search ecosystem revitalized. The 17-18% market share represents the first time a competitor has consistently held a double-digit piece of the pie in over twenty years.

For Alphabet Inc., the response has been aggressive. The recent deployment of Gemini 3 into Google Search marks a „code red“ effort to reclaim the conversational throne. Gemini 3 Flash and Gemini 3 Pro now power „AI Overviews“ that occupy the top of nearly every search result page. However, the competitive advantage currently leans toward ChatGPT in terms of deep engagement. Data from late 2025 indicates that ChatGPT Search users average a 13-minute session duration, compared to Google’s 6-minute average. This „sticky“ behavior suggests that users are not just searching; they are staying to refine, draft, and collaborate with the AI, a level of engagement that traditional search engines have struggled to replicate.

The Wider Significance: The Death of SEO as We Knew It

The broader AI landscape is currently grappling with the „Zero-Click“ reality. With over 65% of searches now being resolved directly on the search results page via AI synthesis, the traditional web economy—built on ad impressions and click-through rates—is facing an existential crisis. This has led to the birth of Generative Engine Optimization (GEO). Instead of optimizing for keywords to appear in a list of links, publishers and brands are now competing to be the cited source within an AI’s conversational answer.

This shift has raised significant concerns regarding publisher revenue and the „cannibalization“ of the open web. While OpenAI and Google have both struck licensing deals with major media conglomerates, smaller independent creators are finding it harder to drive traffic. Comparison to previous milestones, such as the shift from desktop to mobile search in the early 2010s, suggests that while the medium has changed, the underlying struggle for visibility remains. However, the 2026 search landscape is unique because the AI is no longer a middleman; it is increasingly the destination itself.

The Horizon: Agentic Search and Personalization

Looking ahead to the remainder of 2026 and into 2027, the industry is moving toward „Agentic Search.“ Experts predict that the next phase of ChatGPT Search will involve the AI not just finding information, but acting upon it. This could include the AI booking a multi-leg flight itinerary or managing a user’s calendar based on a simple conversational prompt. The challenge that remains is one of privacy and „data silos.“ As search engines become more personalized, the amount of private user data they require to function effectively increases, leading to potential regulatory hurdles in the EU and North America.

Furthermore, we expect to see the integration of multi-modal search become the standard. By the end of 2026, users will likely be able to point their AR glasses at a complex mechanical engine and ask their search agent to „show me the tutorial for fixing this specific valve,“ with the AI pulling real-time data and overlaying instructions. The competition between Gemini 3 and the GPT-5 series will likely center on which model can process these multi-modal inputs with the lowest latency and highest accuracy.

The New Standard for Digital Discovery

The start of 2026 has confirmed that the „Search Wars“ are back, and the stakes have never been higher. ChatGPT’s 17-18% market share is not just a number; it is a testament to a fundamental change in human behavior. We have moved from a world where we „Google it“ to a world where we „Ask it.“ While Google’s 80% dominance is still formidable, the deployment of Gemini 3 shows that the search giant is no longer leading by default, but is instead in a high-stakes race to adapt to an AI-first world.

The key takeaway for 2026 is the emergence of a „duopoly of intent.“ Google remains the primary tool for the physical and commercial world, while ChatGPT has become the primary tool for the intellectual and creative world. In the coming months, the industry will be watching closely to see if Gemini 3 can bridge this gap, or if ChatGPT’s deep user engagement will continue to erode Google’s once-impenetrable fortress. One thing is certain: the era of the „10 blue links“ is officially a relic of the past.

Source: https://markets.financialcontent.com/wral/article/tokenring-2026-1-2-the-search-wars-of-2026-chatgpts-conversational-surge-challenges-googles-decades-long-hegemony

The EV Battery Tech That’s Worth the Hype, According to Experts

Kommentar verfassen

Major battery breakthroughs seemingly happen every day, but only some of that tech ever leaves the lab. WIRED breaks down what’s actually going to change EVs and what’s just a dream.

The EV Battery Tech Thats Worth the Hype According to Experts

You’ve seen the headlines: This battery breakthrough is going to change the electric vehicle forever. And then … silence. You head to the local showroom, and the cars all kind of look and feel the same.

WIRED got annoyed about this phenomenon. So we talked to battery technology experts about what’s really going on in electric vehicle batteries. Which technologies are here? Which will be, probably, but aren’t yet, so don’t hold your breath? What’s probably not coming anytime soon?

“It’s easy to get excited about these things, because batteries are so complex,” says Pranav Jaswani, a technology analyst at IDTechEx, a market intelligence firm. “Many little things are going to have such a big effect.” That’s why so many companies, including automakers, their suppliers, and battery-makers, are experimenting with so many bit parts of the battery. Swap one electrical conductor material for another, and an electric vehicle battery’s range might increase by 50 miles. Rejigger how battery packs are put together, and an automaker might bring down manufacturing costs enough to give consumers a break on the sales lot.

Still, experts say, it can take a long time to get even small tweaks into production cars—sometimes 10 years or more. “Obviously, we want to make sure that whatever we put in an EV works well and it passes safety standards,” says Evelina Stoikou, who leads the battery technology and supply chain team at BloombergNEF, a research firm. Ensuring that means scientists coming up with new ideas, and suppliers figuring out how to execute them; the automakers, in turn, rigorously test each iteration. All the while, everyone’s asking the most important question: Does this improvement make financial sense?

So it’s only logical that not every breakthrough in the lab makes it to the road. Here are the ones that really count—and the ones that haven’t quite panned out, at least so far.

It’s Really Happening

The big deal battery breakthroughs all have something in common: They’re related to the lithium-ion battery. Other battery chemistries are out there—more on them later—but in the next decade, it’s going to be hard to catch up with the dominant battery form. “Lithium-ion is already very mature,” says Stoikou. Lots of players have invested big money in the technology, so “any new one is going to have to compete with the status quo.”

Lithium Iron Phosphate

Why it’s exciting: LFP batteries use iron and phosphate instead of pricier and harder-to-source nickel and cobalt, which are found in conventional lithium-ion batteries. They’re also more stable and slower to degrade after multiple charges. The upshot: LFP batteries can help bring down the cost of manufacturing an EV, an especially important data point while Western electrics struggle to compete, cost-wise, with conventional gas-powered cars. LFP batteries are already common in China, and they’re set to become more popular in European and American electric vehicles in the coming years.

Why it’s hard: LFP is less energy dense than alternatives, meaning you can’t pack as much charge—or range—into each battery.

More Nickel

Why it’s exciting: The increased nickel content in lithium nickel manganese cobalt batteries ups the energy density, meaning more range in a battery pack without much more size or weight. Also, more nickel can mean less cobalt, a metal that’s both expensive and ethically dubious to obtain.

Why it’s hard: Batteries with higher nickel content are potentially less stable, which means they carry a higher risk of cracking or thermal runaway—fires. This means battery-makers experimenting with different nickel content have to spend more time and energy on the careful design of their products. That extra fussiness means more expense. For this reason, expect to see more nickel use in batteries for higher-end EVs.

Dry Electrode Process

Why it’s exciting: Usually, battery electrodes are made by mixing materials into a solvent slurry, which then is applied to a metal current collector foil, dried, and pressed. The dry electrode process cuts down on the solvents by mixing the materials in dry powder form before application and lamination. Less solvent means fewer environmental and health and safety concerns. And getting rid of the drying process can save production time—and up efficiency—while reducing the physical footprint needed to manufacture batteries. This all can lead to cheaper manufacturing, “which should trickle down to make a cheaper car,” says Jaswani. Tesla has already incorporated a dry anode process into its battery-making. (The anode is the negative electrode that stores lithium ions while a battery is charging.) LG and Samsung SGI are also working on pilot production lines.

Why it’s hard: Using dry powders can be more technically complicated.

Cell-to-Pack

Why it’s exciting: In your standard electric vehicle battery, individual battery cells get grouped into modules, which are then assembled into packs. Not so in cell-to-pack, which puts cells directly into a pack structure without the middle module step. This lets battery-makers fit more battery into the same space, and can lead to some 50 additional miles of range and higher top speeds, says Jaswani. It also brings down manufacturing costs, savings that can be passed down to the car buyer. Big-time automakers including Tesla and BYD, plus Chinese battery giant CATL, are already using the tech.

Why it’s hard: Without modules, it can be harder to control thermal runaway and maintain the battery pack’s structure. Plus, cell-to-pack makes replacing a faulty battery cell much harder, which means smaller flaws can require opening or even replacing the entire pack.

Silicon Anodes

Why it’s exciting: Lithium-ion batteries have graphite anodes. Adding silicon to the mix, though, could have huge upsides: more energy storage (meaning longer driving ranges) and faster charging, potentially down to a blazing six to 10 minutes to top up. Tesla already mixes a bit of silicon into its graphite anodes, and other automakers—Mercedes-Benz, General Motors—say they’re getting close to mass production.

Why it’s hard: Silicon alloyed with lithium expands and contracts as it goes through the charging and discharging cycle, which can cause mechanical stress and even fracturing. Over time, this can lead to more dramatic battery capacity losses. For now, you’re more likely to find silicon anodes in smaller batteries, like those in phones or even motorcycles.

It’s Kind of Happening

The battery tech in the more speculative bucket has undergone plenty of testing. But it’s still not quite at a place where most manufacturers are building production lines and putting it into cars.

Sodium-Ion Batteries

Why it’s exciting: Sodium—it’s everywhere! Compared to lithium, the element is cheaper and easier to find and process, which means tracking down the materials to build sodium-ion batteries could give automakers a supply chain break. The batteries also seem to perform better in extreme temperatures, and are more stable. Chinese battery-maker CATL says it will start mass production of the batteries next year and that the batteries could eventually cover 40 percent of the Chinese passenger-vehicle market.

Why it’s hard: Sodium ions are heavier than their lithium counterparts, so they generally store less energy per battery pack. That could make them a better fit for battery storage than for vehicles. It’s also early days for this tech, which means fewer suppliers and fewer time-tested manufacturing processes.

Solid State Batteries

Why it’s exciting: Automakers have been promising for years that groundbreaking solid state batteries are right around the corner. That would be great, if true. This tech subs the liquid or gel electrolytes in a conventional li-ion battery for a solid electrolyte. These electrolytes should come in different chemistries, but they all have some big advantages: more energy density, faster charging, more durability, fewer safety risks (no liquid electrolyte means no leaks). Toyota says it will finally launch its first vehicles with solid state batteries in 2027 or 2028. BloombergNEF projects that by 2035, solid state batteries will account for 10 percent of EV and storage production.

Why it’s hard: Some solid electrolytes have a hard time at low temperatures. The biggest issues, however, have to do with manufacturing. Putting together these new batteries requires new equipment. It’s really hard to build defect-free layers of electrolyte. And the industry hasn’t come to an agreement about which solid electrolyte to use, which makes it hard to create supply chains.

Maybe It’ll Happen

Good ideas don’t always make a ton of sense in the real world.

Wireless Charging

Why it’s exciting: Park your car, get out, and have it charge up while you wait—no plugs required. Wireless charging could be the peak of convenience, and some automakers insist it’s coming. Porsche, for example, is showing off a prototype, with plans to roll out the real thing next year.

Why it’s hard: The issue, says Jaswani, is that the tech underlying the chargers we have right now works perfectly well and is much cheaper to install. He expects that eventually, wireless charging will show up in some restricted use cases—maybe in buses, for example, that could charge up throughout their routes if they stop on top of a charging pad. But this tech may never go truly mainstream, he says.

Source: https://www.wired.com/story/the-ev-battery-tech-thats-worth-the-hype-according-to-experts/

AI agents are starting to eat SaaS and Cloud Software Companies

Kommentar verfassen

Overview

  • Martin Alderson argues that AI coding agents are fundamentally reshaping the build-versus-buy calculus for software, enabling organizations with technical capability to rapidly create custom internal tools that threaten to replace simpler SaaS products—particularly back-office CRUD applications and basic analytics dashboards.
  • Organizations are now questioning SaaS renewal quotes with double-digit annual price increases and choosing to build alternatives with AI agents, while others reduce user licenses by up to 80% by creating internal dashboards that bypass the need for vendor platforms.
  • The disruption poses an acute threat to SaaS companies whose valuations depend on net revenue retention above 100%—a metric that has declined from 109% in 2021 to a median of 101-106% in 2025—as back-office tools now face competition from „engineers at your customers with a spare Friday afternoon with an agent“.​

AI agents are starting to eat SaaS

December 15, 2025·Martin Alderson

We spent fifteen years watching software eat the world. Entire industries got swallowed by software – retail, media, finance – you name it, there has been incredible disruption over the past couple of decades with a proliferation of SaaS tooling. This has led to a huge swath of SaaS companies – valued, collectively, in the trillions.

In my last post debating if the cost of software has dropped 90% with AI coding agents I mainly looked at the supply side of the market. What will happen to demand for SaaS tooling if this hypothesis plays out? I’ve been thinking a lot about these second and third order effects of the changes in software engineering.

The calculus on build vs buy is starting to change. Software ate the world. Agents are going to eat SaaS.

The signals I’m seeing

The obvious place to start is simply demand starting to evaporate – especially for „simpler“ SaaS tools. I’m sure many software engineers have started to realise this – many things I’d think to find a freemium or paid service for I can get an agent to often solve in a few minutes, exactly the way I want it. The interesting thing is I didn’t even notice the shift. It just happened.

If I want an internal dashboard, I don’t even think that Retool or similar would make it easier. I just build the dashboard. If I need to re-encode videos as part of a media ingest process, I just get Claude Code to write a robust wrapper round ffmpeg – and not incur all the cost (and speed) of sending the raw files to a separate service, hitting tier limits or trying to fit another API’s mental model in my head.

This is even more pronounced for less pure software development tasks. For example, I’ve had Gemini 3 produce really high quality UI/UX mockups and wireframes in minutes – not needing to use a separate service or find some templates to start with. Equally, when I want to do a presentation, I don’t need to use a platform to make my slides look nice – I just get Claude Code to export my markdown into a nicely designed PDF.

The other, potentially more impactful, shift I’m starting to see is people really questioning renewal quotes from larger „enterprise“ SaaS companies. While this is very early, I believe this is a really important emerging behaviour. I’ve seen a few examples now where SaaS vendor X sends through their usual annual double-digit % increase in price, and now teams are starting to ask „do we actually need to pay this, or could we just build what we need ourselves?“. A year ago that would be a hypothetical question at best with a quick ’no‘ conclusion. Now it’s a real option people are putting real effort into thinking through.

Finally, most SaaS products contain many features that many customers don’t need or use. A lot of the complexity in SaaS product engineering is managing that – which evaporates overnight when you have only one customer (your organisation). And equally, this customer has complete control of the roadmap when it is the same person. No more hoping that the SaaS vendor prioritises your requests over other customers.

The maintenance objection

The key objection to this is „who maintains these apps?“. Which is a genuine, correct objection to have. Software has bugs to fix, scale problems to solve, security issues to patch and that isn’t changing.

I think firstly it’s important to point out that a lot of SaaS is poorly maintained (and in my experience, often the more expensive it is, the poorer the quality). Often, the security risk comes from having an external third party itself needing to connect and interface with internal data. If you can just move this all behind your existing VPN or access solution, you suddenly reduce your organisation’s attack surface dramatically.

On top of this, agents themselves lower maintenance cost dramatically. Some of the most painful maintenance tasks I’ve had – updating from deprecated libraries to another one with more support – are made significantly easier with agents, especially in statically typed programming ecosystems. Additionally, the biggest hesitancy with companies building internal tools is having one person know everything about it – and if they leave, all the internal knowledge goes. Agents don’t leave. And with a well thought through AGENTS.md file, they can explain the codebase to anyone in the future.

Finally, SaaS comes with maintenance problems too. A recent flashpoint I’ve seen this month from a friend is a SaaS company deciding to deprecate their existing API endpoints and move to another set of APIs, which don’t have all the same methods available. As this is an essential system, this is a huge issue and requires an enormous amount of resource to update, test and rollout the affected integrations.

I’m not suggesting that SMEs with no real software knowledge are going to suddenly replace their entire SaaS suite. What I do think is starting to happen is that organisations with some level of tech capability and understanding are going to think even more critically at their SaaS procurement and vendor lifecycle.

The economics problem for SaaS

SaaS valuations are built on two key assumptions: fast customer growth and high NRR (often exceeding 100%).

I think we can start to see a world already where demand from new customers for certain segments of tooling and apps begins to decline. That’s a problem, and will cause an increase in the sales and marketing expenditure of these companies.

However, the more insidious one is net revenue retention (NRR) declines. NRR is a measure of how much existing customers spend with you on an ongoing basis, adjusted for churn. If your NRR is at 100%, your existing cohort of customers are spending the same. If it’s less than that then they are spending less with you and/or customers are leaving overall.

Many great SaaS companies have NRR significantly above 100%. This is the beauty of a lot of SaaS business models – companies grow and require more users added to their plan. Or they need to upgrade from a lower priced tier to a higher one to gain additional features. These increases are generally very profitable. You don’t need to spend a fortune on sales and marketing to get this uptick (you already have a relationship with them) and the profit margin of adding another 100 user licenses to a SaaS product for a customer is somewhere close to infinity.

This is where I think some SaaS companies will get badly hit. People will start migrating parts of the solution away to self-built/modified internal platforms to avoid having to pay significantly more for the next pricing tier up. Or they’ll ingest the data from your platform via your APIs and build internal dashboards and reporting which means they can remove 80% of their user licenses.

Where this doesn’t work (and what still has a moat)

The obvious one is anything that requires very high uptime and SLAs. Getting to four or five 9s is really hard, and building high availability systems gets really difficult – and it’s very easy to shoot yourself in the foot building them. As such, things like payment processing and other core infrastructure are pretty safe in my eyes. You’re not (yet) going to replace Stripe and all their engineering work on core payments easily with an agent.

Equally, very high volume systems and data lakes are difficult to replace. It’s not trivial to spin up clusters for huge datasets or transaction volumes. This again requires specialised knowledge that is likely to be in short supply at your organisation, if it exists at all.

The other one is software with significant network effects – where you collaborate with people, especially external to your organisation. Slack is a great example – it’s not something you are going to replace with an in-house tool. Equally, products with rich integration ecosystems and plugin marketplaces have a real advantage here.

And companies that have proprietary datasets are still very valuable. Financial data, sales intelligence and the like stay valuable. If anything, I think these companies have a real edge as agents can leverage this data in new ways – they get more locked in.

And finally, regulation and compliance is still very important. Many industries require regulatory compliance – this isn’t going to change overnight.

This does require your organisation having the skills (internally or externally) to manage these newly created apps. I think products and people involved in SRE and DevOps are going to have a real upswing in demand. I suspect we’ll see entirely new functions and teams in companies solely dedicated to managing these new applications. This does of course have a cost, but this cost can be often managed by existing SRE or DevOps functions, or if it requires new headcount and infrastructure, amortised over a much higher number of apps.

Who’s most at risk?

To me the companies that are at serious risk are back-office tools that are really just CRUD logic – or simple dashboards and analytics on top of their customers‘ own data.

These tools often generate a lot of friction – because they don’t work exactly the way the customer wants them to – and they are tools that are the most easily replaced with agents. It’s very easy to document the existing system and tell the agent to build something, but with the pain points removed.

SaaS certainly isn’t dead. Like any major shifts in technology, there are winners and losers. I do think the bar is going to be much higher for many SaaS products that don’t have a clear moat or proprietary knowledge.

What’s going to be difficult to predict is how quickly agents can move up the value chain. I’m assuming that agents can’t manage complex database clusters – but I’m not sure that’s going to be the case for much longer.

And I’m not seeing a path for every company to suddenly replace all their SaaS spend. If anything, I think we’ll see (another) splintering in the market. Companies with strong internal technical ability vs those that don’t. This becomes yet another competitive advantage for those that do – and those that don’t will likely see dramatically increased costs as SaaS providers try and recoup some of the lost sales from the first group to the second who are less able to switch away.

But my key takeaway would be that if your product is just a SQL wrapper on a billing system, you now have thousands of competitors: engineers at your customers with a spare Friday afternoon with an agent.

Source: https://martinalderson.com/posts/ai-agents-are-starting-to-eat-saas/

The Privacy-Friendly Tech to Replace Your US-Based Email, Browser, and Search

Kommentar verfassen

Thanks to drastic policy changes in the US and Big Tech’s embrace of the second Trump administration, many people are moving their digital lives abroad. Here are a few options to get you started.

Image may contain Electronics Screen Computer Hardware Hardware and Monitor

From your email to your web browsing, it’s highly likely that your daily online life is dominated by a small number of tech giants—namely Google, Microsoft, and Apple. But since Big Tech has been cozying up to the second Trump administration, which has taken an aggressive stance on foreign policy, and Elon Musk’s so-called Department of Government Efficiency (DOGE) has ravaged through the government, some attitudes towards using US-based digital services have been changing.

While movements to shift from US digital services aren’t new, they’ve intensified in recent months. Companies in Europe have started moving away from some US cloud giants in favor of services that handle data locally, and there have been efforts from officials in Europe to shift to homegrown tech that has fewer perceived risks. For example, the French and German governments have created their own Docs word processor to rival Google Docs.

Meanwhile, one consumer poll released in March had 62 percent of people from nine European countries saying that large US tech companies were a threat to the continent’s sovereignty. At the same time, lists of non-US tech alternatives and European-based tech options have seen a surge in visitors in recent months.

For three of the most widely used tech services—email, web browsers, and search engines—we’ve been through some of the alternatives that are privacy-focused and picked some options you may want to consider. Other options are available, but these organizations and companies aim to minimize data they collect and often put privacy first.

There are caveats, though. While many of the services on this list are based outside of the US, there’s still the potential that some of them rely upon Big Tech services themselves—for instance, some search engines can use results or indexes provided by Big Tech, while companies may use software or services, such as cloud hosting, that are created by US tech firms. So trying to distance yourself entirely may not be as straightforward as it first looks.

Web Browsers

Mullvad

Based in Sweden, Mullvad is perhaps best known for its VPN, but in 2023 the organization teamed up with digital anonymity service Tor to create the Mullvad Browser. The open source browser, which is available only on desktop, says it collects no user data and is focused on privacy. The browser has been designed to stop people from tracking you via browser fingerprinting as you move around the web, plus it has a “private mode” that isolates tracking cookies enabled by default. “The underlying policy of Mullvad is that we never store any activity logs of any kind,” its privacy policy says. The browser is designed to work with Mullvad’s VPN but is also compatible with any VPN that you might use.

Vivaldi

WIRED’s Scott Gilbertson swears by Vivaldi and has called it the web’s best browser. Available on desktop and mobile, the Norwegian-headquartered browser says it doesn’t profile your behavior. “The sites you visit, what you type in the browser, your downloads, we have no access to that data,” the company says. “It either stays on your local machine or gets encrypted.” It also blocks trackers and hosts data in Iceland, which has strong data protection laws. Its privacy policy says it anonymizes IP addresses and doesn’t share browsing data.

Search Engines

Qwant French search engine Qwant has built its own search index, crawling more than 20 billion pages to create its own records of the web. Creating a search index is a hugely costly, laborious process, and as a result, many alternative search engines will not create an extensive index and instead use search results from Google or Microsoft’s Bing—enhancing them with their own data and algorithms. Qwant says it uses Bing to “supplement” search results that it hasn’t indexed. Beyond this, Qwant says it does not use targeted advertising, or store people’s search history. “Your data remains confidential, and the processing of your data remains the same,” the company says in its privacy policy.

Mojeek

Mojeek, based out of the United Kingdom, has built its own web crawler and index, saying that its search results are “100% independent.” The search engine does not track you, it says in its privacy policy, and only keeps some specific logs of information. “Mojeek removes any possibility of tracking or identifying any particular user,” its privacy policy says. It uses its own algorithms to rank search results, not using click or personalization data to create ranks, and says that this can mean two people searching for the same thing while in different countries can receive the same search results.

Startpage

Based in the Netherlands, Startpage says that when you make a search request, the first thing that happens is it removes your IP address and personal data—it doesn’t use any tracking cookies, it says. The company uses Google and Bing to provide its search results but says it acts as an “intermediary” between you and the providers. “Startpage submits your query to Google and Bing anonymously on your behalf, then returns the results to you, privately,” it says on its website. “Google and Microsoft do not know who made the search request—instead, they only see Startpage.”

Ecosia

Nonprofit search engine Ecosia uses the money it makes to help plant trees. The company also offers various privacy promises when you search with it, too. Based in Germany, the company says it doesn’t collect excessive data and doesn’t use search data to personalize ads. Like other search alternatives, Ecosia uses Google’s and Bing’s search results (you can pick which one in the settings). “We only collect and process data that is necessary to provide you with the best search results (which includes your IP address, search terms and session behavioral data),” the company says on its website. The information it collects is gathered to provide search results from its Big Tech partners and detect fraud, it says. (At the end of 2024, Ecosia partnered with Qwant to build more search engine infrastructure in Europe).

Email Providers

ProtonMail

Based in Switzerland, Proton started with a privacy-focused email service and has built out a series of apps, including cloud storage, docs, and a VPN to rival Google. The company says it cannot read any messages in people’s inboxes, and it offers end-to-end encryption for emails sent to other Proton Mail addresses, as well as a way to send password protected emails to non Proton accounts. It blocks trackers in emails and has multiple account options, including both free and paid choices. Its privacy policy describes what information the company has access to, which includes sender and recipient email addresses, plus IP addresses where messages arrive from, message subject lines, and when emails are sent. (Despite Switzerland’s strong privacy laws, the government has recently announced it may require encrypted services to keep user’s data, something that Proton has pushed back on).

Tuta

Tuta, which used to be called Tutanota and is based in Germany, says it encrypts email content, subject lines, calendars, address books, and other data in your inbox. “The only unencrypted data are mail addresses of users as well as senders and recipients of emails,” it says on its website, adding that users‘ encryption keys cannot be accessed by developers. Like Proton, emails sent between Tuta accounts are end-to-end encrypted, and you can send password protected emails when messaging an account from another email provider. The company also has an end-to-end encrypted calendar and offers both free and paid plans.

Source: https://www.wired.com/story/the-privacy-friendly-tech-to-replace-your-us-based-email-browser-and-search/