The release of Claude Fable 5 — a 10-trillion-parameter model that compressed months of Stripe engineering into a single day — crystallized a week defined by two colliding forces: extraordinary capability gains at the frontier and deep structural uncertainty about who captures the value. Benedict Evans's framework for why foundation models resemble telecom infrastructure more than operating systems ran directly into real-world evidence of explosive demand, while a new engineering paradigm — loop design over prompt engineering — quietly separated the top 0.3% of AI practitioners from everyone else.
Foundation models look like telecom infrastructure, not operating systems — and that's the bear case
The Economics of AI Usage and What's Next For SaaS · Benedict Evans on a16z
Benedict Evans's core thesis: foundation models lack the levers that let platform layers capture value. Unlike Windows or iOS, they have no network effects, no control up the stack, and no lock-in — enterprises often don't even know which model their SaaS vendor uses.
The mobile analogy is damning: data traffic rose 1,500–2,000x, telcos earned $1 trillion in revenue, spent $200B/year on capex — and their stocks were flat for 20 years
Microsoft, Meta, and Google are guiding $700B in capex this year, spending on infrastructure — more than telecoms, approaching oil and gas
Source episodes
Sourced from 78 episodes across 11 podcasts this week
With 3–6 frontier model companies competing, some cross-subsidized by ad revenue (Google), price discipline is structurally impossible
Evans's committed prediction: "I don't think foundation models are a product. I don't think a chatbot is a product. I think the value will be further up."
"the foundation models seem to look more like that. They seem to more look more like the hyperscalers in that sense in that they don't have you know they might have competitive advantages that further up the stack you don't have leverage you don't have a network effect you don't have control."
Jevons paradox is already playing out in AI infrastructure — cheaper models grow the market
Nebius Co-Founder on AI Infrastructure Bubbles · How Price Elastic is Demand for Compute
The bear case that cheaper inference destroys infrastructure demand is empirically wrong. Nebius's own data from DeepSeek week is the sharpest available counterexample.
Nebius stock fell 40% during DeepSeek week while the company had its best-ever sales week
Every time a unit of intelligence gets cheaper, consumption increases — users solve more complex tasks with the same budget
Open-source models are already replacing closed models at scale: Revolut moved from 99% OpenAI spend to open-source after building evaluation and CI/CD infrastructure
The "cold start problem": enterprises look slow until they build eval foundations, then grow exponentially
"Every time we got intelligence cheaper uh this the same unit of intelligence cheaper, we are not reducing the consumption but we increasing the consumption because we can just solve more complex tasks with the same budget"
Claude Fable 5 is a 10-trillion-parameter model with a 2x lead on the hardest coding benchmarks
MYTHOS is LIVE!!!! (2026-06-09)
Anthropic's Fable 5 — internally known as Mythos — is the first publicly available 10-trillion-parameter model, and its benchmark lead widens on harder tasks.
SWEBench Pro: Fable 80% vs. Opus 4.8 at 69% vs. GPT-5.5 at 58%
Frontier Code Diamond: Fable 29.3% vs. Opus 4.8 at roughly half that, GPT-5.5 at 5.7%
Stripe compressed months of engineering into days on a 50-million-line Ruby codebase; a whole-team, two-month migration done in one day
Pricing: $10/M input, $50/M output — less than half the Mythos preview price
Host's updated assessment: open-source models are now ~1 year behind frontier, up from a prior estimate of 6 months
"Stripe reported that Fable 5 compressed months of engineering into days in a 50 millionline Ruby codebase. The model performed a codebasewide migration in a day that would otherwise have taken a whole team over two months by hand."
Anthropic withheld Fable for ~5 months and built a silent distillation countermeasure
MYTHOS is LIVE!!!! (2026-06-09) · Mythos is officially here. (2026-06-09)
Anthropic tested Mythos internally since January 2026, announced it publicly months before release, and only shipped June 9 — likely to use it to accelerate their own next model before competitors could distill it.
Requests flagged as distillation attempts are silently redirected to Opus 4.8 — the attacker gets the older model without knowing it
The model was released publicly only after additional safety guardrails were added; internally it was framed as potentially dangerous
There appears to be no ceiling on quality gains from more thinking tokens — Noam Brown (OpenAI) published a blog post showing models keep improving even at 100M thinking tokens, with major implications for inference demand
"Requests that are flagged by our classifiers as being part of such distillation attempts will fall back to opus 4.8. They're like, 'Oh, you want to distill? Go ahead. You can have our old model.'"
The new engineering meta: design loops, not prompts — but only 1-in-2,500 practitioners are there
Only the best are using them... (2026-06-09)
Two of the most prominent AI coding practitioners — one from OpenAI, one from Anthropic — independently went viral with the same message: stop prompting agents, start designing loops.
A loop requires exactly two things: a trigger (event-based, scheduled, or human-initiated) and a verifiable goal — structurally identical to reinforcement learning
The critical distinction: a loop decides if it reached its goal; an automation just executes steps
Claude Code has a native /loop command built in — immediately actionable
Non-deterministic goals are the core failure mode: without a clean end state, agents burn tokens indefinitely
Peter Steinberger's monthly token bill: $1.3 million — the technique is currently capital-gated to a tiny elite
Only 1-in-2,500 AI users actually use agents in any capacity; 84% of the world has never used AI at all
"I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and kind of figuring out what to do. My job is to write loops."
Karpathy writes zero lines of code; the coding TAM alone is half a trillion dollars
Why the AI Boom Is Just Getting Started (2026-06-09)
Andrej Karpathy's flip — from writing 80% of code by hand to writing zero lines except in English — is the primary-source signal that a genuine phase transition has occurred, not incremental improvement.
Bottom-up TAM math: 20 million coders × $20–30K/year in token spend ≈ $500B — derived from observing Anthropic employees spending $100/day on tokens
Anthropic investment entered at $1.8B valuation; the firm expected $9B, got far more
Anthropic already has half the compute it needs — before the mainstream adoption wave; Marc Andreessen predicts insufficient compute for four years
Sundar Pichai estimates only 10 basis points of knowledge workers are truly using AI — the S-curve hasn't started
Recursive improvement flywheel: Anthropic feeds Claude-written code back into model training, accelerating innovation in a self-reinforcing loop
"Karpathy said last year's code tools could write 20% and 80% would be handwritten — that flipped when the latest model came out and now he hasn't written a line of code except in English"
SaaS trades at a discount to the S&P 500 for the first time in history
SpaceX Launches Largest Ever IPO | OpenAI Files to Go Public | Uber Cuts 23% of HR (2026-06-11)
The asset class that commanded premium multiples for a decade has been repriced below the broad market — a historic inflection point reflecting AI disruption fears and slowing growth.
Evans's framework explains the mechanism: AI productivity gains become competitive necessities competed away — if a DCF takes 10 seconds instead of a week, you do 50 DCFs but can't charge more
Four compounding headwinds for legacy software (per Why the AI Boom Is Just Getting Started): lower CIO priority, budget pressure from AI token spend, loss of annual pricing power, and seat reduction from AI
One firm sold almost all enterprise software and went net short entering 2025 — the position paid off in Q1
Evans's heuristic: LLMs excel where "the average answer" is what you want — tasks requiring novel judgment or unexplained expertise are not automatable
"SaaS now trades at a discount to the S&P 500 for the first time in history."
Meta's Biohub claims no other organization combines frontier AI and frontier biology — and has the results to back it
"Curing All Disease by next century is too conservative" - Mark Zuckerberg (2026-06-10)
Zuckerberg's original goal of curing all disease by 2100 now seems too conservative given AI progress in biology — a striking reversal from a goal that Nobel Prize-winning scientists laughed at.
ESM Fold predicted structures for over 1.1 billion proteins; antibody design emerged as an emergent property of a general protein model, not a specialized one
Digital screening narrows millions of candidates to a 96-well plate with nanomolar binders — a massive compression of experimental cycles
Single-cell atlas can predict off-target toxicity before human trials by revealing which unexpected cell types express a drug's target receptor
Core constraint: biology data doesn't exist yet — it must be invented, not collected, unlike internet text for LLMs
"I don't actually think that there's any other organization in the world that's doing both the frontier biology and the frontier AI."
NVIDIA is building the end-to-end physical AI stack for healthcare — the autonomous vehicle playbook, applied to surgery
NVIDIA GTC 2026: Healthcare Special Address (2026-06-09)
NVIDIA is stacking Open-H data, Groot-H vision-language-action models, Cosmos-H simulation, and Project Rio hospital digital twins into a full-stack healthcare robotics platform — the same moat strategy it used in autonomous vehicles.
Open-H: world's first surgical robotics dataset — 750 hours of video + kinematics, 11 embodiments, 35+ global contributors
Groot-H: first generalizable VLA model for surgical robotics — image in, language prompt in, action out
Healthcare robotics must navigate 160,000 unique hospitals and 72,000 medical procedures — making synthetic data and simulation essential
Roche is tripling its AI factory footprint to over 3,500 Blackwell GPUs across US and European sites
85% of AI healthcare revenue flows to startups, not incumbents — over 2,000 of NVIDIA's 5,000 Inception companies are in digital health
AlphaFold 2 generated 200 million protein structures from a base of 200,000 — the clearest articulation of "compute is data"
"we've essentially assembled the first and only physical AI end-to-end platform for healthcare robotics domain specific. So, we're taking the amazing breakthroughs of self-driving cars. We're taking the amazing breakthroughs of generalized robotics, and we're helping specialize it."
The GPT-5.5 parameter estimate paper: viral headline, shaky methodology
Paper Club: Incompressible Knowledge and Model Fingerprints (2026-06-11)
The paper's central claim — that GPT-5.5 has ~9.7 trillion parameters — went viral, but a single code bug (a flooring operation) collapses the estimate from 8.8T to 1.9T.
Core theoretical contribution is solid: factual bits cannot be compressed regardless of architecture — a Shannon entropy floor exists for storing specific facts
A 2026 7B model matches a 2023 70B model on MMLU — but this reflects compression of procedural/linguistic parameters, not factual ones
IK score increases only 14.7% for every 10x parameter increase — factual knowledge accumulates logarithmically
Every frontier model scores ~0% on T7, the hardest factual tier — significant headroom remains
The 90% confidence interval on GPT-5.5 spans 3.2T to 29T — a nearly 10x range that most early readers ignored
"9.7 with 90% confidence from 3.2 trillion to 29 trillion, which is quite a range."
Key Takeaways
Foundation models are infrastructure, not platforms. Benedict Evans's mobile analogy is the sharpest available framework: massive capex, enormous usage growth, and value migrating up the stack to someone else — the same pattern that left telcos with flat stocks for 20 years.
Fable 5 is a genuine capability step-change, not a benchmark artifact. A 2x lead on Frontier Code Diamond, Stripe compressing months of engineering into a day, and Karpathy writing zero lines of code are converging signals that the frontier has moved.
Loop engineering is the new elite skill — and it's capital-gated. The shift from prompting to designing verifiable agent loops is real and consequential, but only 1-in-2,500 AI users are there, partly because the token bills run into seven figures monthly.
Jevons paradox is the correct mental model for AI infrastructure demand. Cheaper models grow consumption; Nebius's best-ever sales week during DeepSeek's 40% stock drop is the clearest empirical proof available.
Biology is the next frontier for AI capability demonstrations. Meta's Biohub combining frontier AI with frontier wet labs, NVIDIA's full-stack surgical robotics platform, and AlphaFold's "compute is data" flywheel all point to life sciences as the domain where AI's next phase-transition will be most visible.
SaaS's historic discount to the S&P 500 is the market pricing in structural disruption. Four simultaneous headwinds — CIO priority, budget pressure, pricing power, and seat reduction — make the bear case for legacy software more than a valuation story.