TubeReads

Why AI Feels Like the Internet in 1997 | Benedict Evans on a16z

Silicon Valley awaits Benedict Evans' annual presentation with the anticipation usually reserved for product launches. This year, his «AI eats the world» thesis confronts a landscape that has suddenly narrowed: agentic coding went from «kind of useful to really changing everything,» while most other use cases remain frustratingly vague. Yet foundation model companies are raising more capital than ever, even as Evans argues they're destined to become commoditized infrastructure—much like telecom operators who built a trillion-dollar network but captured none of the value. The central tension: are we building the next Windows, or just extremely expensive plumbing?

Durée de la vidéo : 1:00:33·Publié 8 juin 2026·Langue de la vidéo : English
7–8 min de lecture·13,362 mots prononcés → résumé en 1,458 mots (9x)·

1 —

Points clés

1

Agentic coding has achieved genuine product-market fit, pulling itself out of developers' hands, while consumer and enterprise use cases remain weekly habits rather than daily necessities.

2

Foundation models appear to lack network effects, sustainable differentiation, or pricing power—more like telecom infrastructure than operating systems—suggesting value will move up the stack to applications.

3

The current supply-demand-price equilibrium is transitory; with $700 billion in annual capex and models improving 100–200× in efficiency yearly, pricing will eventually align with marginal cost.

4

The hardest problems are not technological but organizational: understanding what work actually consists of, how to reconfigure business models, and what becomes possible when formerly expensive tasks become trivial.

5

History suggests the companies building the infrastructure rarely capture the value—mobile operators spent trillions and saw flat stock prices for 20 years while Google, Meta, and Apple reaped the rewards.

En bref

AI has found undeniable product-market fit in software development, but elsewhere we're still in 1997—excited, but unclear what it's for. The models will likely become commodities; the real value will accrue to whoever figures out what to build on top, and that discovery process will take years, not quarters.


2 —

The Year Everything Narrowed to Code

Agentic coding achieved breakout success while other use cases remain uncertain.

WHAT WORKED
Software Development Finds Product-Market Fit
In the past year, agentic coding went from «kind of useful to really changing everything.» OpenAI's strategy pivoted from «everything all at once» to doubling down on coding after Anthropic demonstrated traction. The result: customers are pulling the product out of developers' hands, and the entire tech industry has narrowed its focus massively onto this single validated use case.
WHAT DIDN'T
Consumer Adoption Remains Weekly, Not Daily
Outside of software development, AI usage patterns reveal a stark divide: valley insiders running clusters of Mac Studios with Claude all day versus 40% of people who «used it last week for something.» Despite 900 million weekly active users across chatbots, no consumer application has achieved the daily necessity that defines true platform adoption. The fundamental questions from two years ago—winner-takes-all dynamics, value capture, sustainable differentiation—remain unanswered.

3 —

The Commoditization Thesis

Foundation models lack network effects and may face the telecom fate.

Evans argues foundation models are structurally positioned to become commoditized infrastructure rather than defensible platforms. Unlike iOS or Windows, models lack network effects, sustainable competitive moats, or levers to control what gets built on top. With three to six frontier model providers competing—some with alternative business models like Google's ad revenue—pricing power seems unlikely to persist once the current supply crunch resolves.

The historical parallel is striking: mobile network operators collectively earn about a trillion dollars in revenue, spend $200 billion yearly on capex, and have seen flat stock prices for two decades while traffic grew 1,500–2,000×. All the value—and all the innovation—moved to companies like Google, Meta, and Apple who built on top of the infrastructure. The operators thought they would build mobile banking and capture the upside; instead, they became expensive plumbing.

The central question: can model providers do «the whole thing,» or will value accrue to 300 specialized applications built on top? If you need dedicated software for tax preparation, legal work, or financial advice—not just a chatbot—then models become the new cloud layer: essential, expensive to build, but sold at marginal cost with no pricing power.


4 —

The Trillion-Dollar Capex Puzzle

Extreme scarcity meets extreme investment in an unsustainable equilibrium.

Big Four AI Capex Guidance (2024)
$700 billion
Microsoft, Meta, Google, and Amazon are each projected to spend over 50% of revenue on capital expenditures—telecoms typically spend 15–20%.
Mobile Network Annual Capex
$200 billion
Global mobile operators' infrastructure spending—less than one-third of projected AI capex.
Model Efficiency Improvement Rate
100–200× per year
Annual gains in computational efficiency mean today's pricing structures are inherently temporary.
Mobile Data Traffic Growth (15 years)
1,500–2,000×
Despite massive usage increases, telecom operator stocks remained flat for 20 years while application layer companies captured value.
Physical Spending Limit
Cannot exceed $10 trillion/year
There are laws-of-physics caps on available capital; current trajectory is unsustainable regardless of ROI.

5 —

What Computers Couldn't Do—Until Now

🔍
Semantic Understanding
Google and Meta knew products were SKUs and metadata, but not what they were or why people bought them. LLMs can understand «show me coats like this but different prices» or «suggest a winter coat that changes my look, but not too much»—queries that were science fiction three years ago.
📊
Cross-System Synthesis
«Look at all our Zoom calls, email flows in Salesforce, product telemetry, and analytics—how should we change pricing to improve churn?» This represents a new layer of abstraction: asking questions that span multiple systems and require judgment, not just aggregation.
💰
Cash Flow Prediction
A commodities company dealing with small producers and uncertain invoice timing can now use LLMs for better cash flow forecasting—a low-margin business where this matters enormously. These point solutions outside tech represent real enterprise adoption.
⚡
Jevons Paradox in Action
If a DCF takes a week, you do two. If it takes 10 seconds, you do 50—but can't charge more. Productivity gains become competitive necessities that get competed away, creating consumer surplus without captured value. Excel's story repeating.

6 —

Where the Questions Move Next

AI problems become industry problems—lawyers, not engineers, must answer them.

Evans argues the most important questions are now leaving Silicon Valley. What does AI mean for law firms with pyramid staffing models when you can automate junior associate work? The answer requires understanding what clients actually pay for, what associates actually do, and how those align—knowledge that exists in AmLaw 100 partnerships, not in San Francisco. Same for consultancies, investment banks, advertising agencies, and the Big Four accounting firms.

This mirrors what happened with Netflix: all the enabling technology came from tech, but all the strategic questions—which shows, how many, what talent costs, movies or sports—were Los Angeles questions. «Content isn't king» meant the value and decisions moved to the content industry. Similarly, generative video's impact on Hollywood is a question for people like Ben Affleck, who built and sold a production company, not for AI researchers.

The shift represents a fundamental change in where expertise matters. Vertical AI startups face the challenge that often the industry doesn't know a problem exists—many a16z investments solved problems that took two years just to explain to customers. The idea that a middle manager will use a general-purpose tool to solve a global industry problem misunderstands how innovation actually works: someone outside the industry spots the hole in the universe, builds the solution, then persuades the industry it was always needed.


7 —

The 1997 Mindset

Predictions are impossible; focus on asking the right questions instead.

“It's going to be magic and in 20 years time we'll just say, well, of course that's how it is. Computers have always done that.”

— Benedict Evans


8 —

Why We Can't Predict What Happens Next

Unlike previous platform shifts, we don't know AI's fundamental physical limits.

💡

Why We Can't Predict What Happens Next

Every previous technology wave had known constraints: in 1995, you knew PCs cost $3,000 and telecoms wouldn't deliver global broadband overnight. With AI, we might check our phones after this recording and find a push notification that a new model costs 2% of the old price because someone «worked something out.» Not knowing the physical limits—how much bigger, better, faster, cheaper models can get—makes this fundamentally different and more unpredictable than the PC, internet, or mobile eras.


9 —

The Software Industry's Uncertain Future

More software, not less—but with radically different margin structures.

1

Fragmented Landscape Enterprises run big horizontal systems (SAP, Workday), 300–400 SaaS apps, 1,000 internal apps, plus a «fuzzy improvised space» of Excel, email, and shared drives. AI arrives into this complexity as another set of options.

2

Bottom or Top of Stack? Does the LLM become a feature inside Salesforce («suggest what to say on this call based on history and objectives») or does it sit on top, synthesizing across Salesforce, Workday, email, and analytics to answer questions no single system could?

3

Build vs. Buy Shifts Building software becomes way cheaper and faster, enabling more competition and more specialized tools. Tasks that lived in Excel might move to dedicated apps—or might now be handled by LLMs directly, creating a new layer in the build-versus-buy decision tree.

4

The SaaS Apocalypse Question Investors know some percentage of existing SaaS companies will be «wiped out,» but can't identify which ones. The rational response: derate the entire sector until the picture clarifies, even if the overall result is more software, not less.


10 —

Titres mentionnés

MSFTMicrosoft
GOOGLAlphabet / Google
METAMeta Platforms
AAPLApple
AMZNAmazon
NVDANvidia

11 —

Personnes

Benedict Evans
Former a16z Partner, Industry Analyst
guest
Marc Andreessen
Co-founder of a16z, former Netscape founder
mentioned
Dario Amodei
CEO of Anthropic
mentioned
Ben Affleck
Actor, Producer
mentioned

Glossaire
Agentic codingAI systems that autonomously write, debug, and modify code with minimal human intervention, moving beyond autocomplete to fundamental workflow transformation.
Foundation modelLarge-scale AI models (like GPT-4, Claude) trained on broad data that serve as the base layer for various applications, analogous to an operating system for AI.
TokenThe unit of text (roughly 0.75 words) that LLMs process and generate; usage and pricing are typically measured in tokens consumed.
Jevons paradoxWhen technological progress makes a resource more efficient, total consumption increases rather than decreases—e.g., cheaper computation leads to far more computation being used.
Network effectWhen a product becomes more valuable as more people use it (e.g., telephones, social networks); a key source of defensibility that AI models appear to lack.

Avertissement : Ceci est un résumé généré par IA d'une vidéo YouTube à des fins éducatives et de référence. Il ne constitue pas un conseil en investissement, financier ou juridique. Vérifiez toujours les informations auprès des sources originales avant de prendre des décisions. TubeReads n'est pas affilié au créateur de contenu.