Summary

Morin’s headline call: the H100 is a bubble that may blow sometime — “probably this year” for the potential compute oversupply/distress scenario. The market runs on the A100 financial model — generation zero trains, last generation does inference — but the H100 arrived at 5x the price with only ~2x the inference performance, financed by six-to-seven-year amortization plans with the GPUs themselves as collateral. “Something has got to give” — he’s already getting cold-email discounts from providers he’s never heard of, and speculates GPUs may be bought for “30 cents on the dollar.”
The value split in five years: “95% inference, 5% training.” Agents and reasoning flip compute economics from throughput-bound to latency-bound — a GPU can serve 10,000 tokens/second across 100 streams but not to one user — and “this is where Nvidia can be attacked.” Yet forced to buy one stock today, he still picks Nvidia “because the supply,” hoping to come back later and say buy AMD “as much as you can.”
Google is the sleeping giant. His winning triangle is products, data, and compute — “who has all three? Google… they can sprinkle everywhere.” The shallow take is that OpenAI threatens search; in reality “OpenAI is amazing, but it’s not their compute” — it’s Microsoft’s, and “if you don’t own your compute, you’re starting with something at your ankle.”
The margin stack is the tradeable insight for anyone building on AI: TSMC sells at 60% margin, Nvidia at 90%, Amazon takes another ~30% — “you are a very thin crust on a very big cake.” His advice to startups: do not resell compute if you can; of $1 of spend, ~98% is somebody else’s margin, impossible to square into a $20/month subscription.
Switching economics explain why AMD hasn’t closed the gap despite ~4x inference efficiency: “being seven times better is not enough to get people to switch — people will choose nothing over something.” The workaround is top-down: Microsoft bought AMD’s supply and runs ChatGPT on it, which puts OpenAI’s inference “in the green.” ZML’s thesis is to make the buy-in zero so “if you are 30% better, I’ll switch to you.”
On new silicon: Gro and Cerebras prove single-stream speed via SRAM but the economics don’t scale (Gro has 230MB per chip; a 70B model in BF16 is 140GB — “you do the math”). The companies he sees going in the direction of lower prices: Etched and likely VSORA — “if you can deliver this at a price comparable to GPUs, you’ve won” — with compute-in-memory (likely Rain AI, Fractile) as the frontier after that.
Stargate: “I don’t buy it… it’s an American car of AI — it’s big, it consumes a lot of gas, but ultimately it’s not a good car.” The real limits are “talent and energy, that’s it.” Meanwhile Blackwell orders are getting cancelled over heat-dissipation and chip-bending problems, and Jensen’s true challenge is “how to navigate the down slope — the highs are very high, but they don’t last forever.”

Deep dive

1. The H100 bubble may blow — GPUs potentially at 30 cents on the dollar

Morin’s core warning: the industry is built on the A100 financial model — generation zero does training, last generation does inference — “and it worked beautifully.” Then the H100 came in at 5x the price with maybe 2x the inference performance (at launch it ran inference at the same speed as the A100). “There’s a money gap that’s going to have to be bridged… something has got to give.”
What worries him most: amortization plans of six, seven years with the GPUs as the collateral. His speculative scenario: a data center somewhere in the US with a thousand GPUs “that people may buy 30 cents on the dollar.” Timeframe: “probably this year.”
The tell that it’s started: “I’m getting cold emails for discounts from services I never heard about” — since roughly October–November. Building a training cluster and building a cloud provider are different businesses, and some people are left with capex they don’t know what to do with.
The same down-slope logic reaches Jensen: Blackwell orders are getting cancelled over heat and chip-bending problems; H100 supply was smoothed across the year, and some buyers still haven’t received last year’s orders while the B200 is already visible — “and they’re super pissed.” “The highs are very high, but they don’t last forever.”

2. Agents and reasoning flip compute from throughput-bound to latency-bound

The mechanism behind his whole thesis: chat rewards streaming throughput, but with agents and reasoning “you need to wait until the end of the request to get whatever it is you came for” — all that matters is time from request to completed answer. GPUs can serve 10,000 tokens/second across a hundred streams, “but they cannot give you 10,000 tokens per second only on you.” His verdict: “this is where I think Nvidia can be attacked.”
Nvidia’s current approach is vertical — “more of more” — and it’s straining physics: Blackwell assembled two chips into a surface so large the chip started to bend, breaking heat-sink contact, at a 1,000-watt envelope requiring liquid cooling.
The next turn of the screw is latent-space reasoning — thinking without translating to tokens, the way “if you think to yourself, you don’t say out loud what you’re thinking.” That plus likely Yann LeCun-style energy-based models “fundamentally GPUs cannot deliver at scale,” because external-memory access is the wall: HBM is the best we have and it’s still very slow next to SRAM.

3. GPUs were never AI chips — Nvidia’s training win was Mellanox

The history he wants remembered: GPGPU was a “cool trick” from ~20 years ago — fooling a graphics renderer into doing parallel math. “GPUs are a good trick for AI, but they’re not built for AI.” TPUs were the pioneering dedicated architecture; for LLMs the GPU trick “starts to crack” under memory transfers.
His revisionist claim: Nvidia won training because of Mellanox, not the compute — a matrix is read “hundreds of times” during multiplication, so InfiniBand interconnect is “the name of the game” in training. “When you do inference — not so much. You don’t care.” Model sizes are literally chosen so people can serve them without connecting machines.
Why the lock-in persists anyway: the PyTorch–CUDA duo is “very, very hard to break,” and a self-perpetuating circle — people buy Nvidia to resell because everyone runs Nvidia — keeps the market saturated with the least efficient option, which he argues isn’t even the best software platform.

4. “You are a very thin crust on a very big cake”

The dirty secret, as he stacks it: TSMC sells at 60% margin, Nvidia at 90%, then Amazon takes another ~30%. Hence his standing advice to AI startups: “do not resell compute if you can” — for $1 of spend, “maybe 98% of it goes to somebody else’s margin,” a very hard circle to square into $20 a month.
On CUDA as a moat: “They spend a lot of energy making you care about stuff you shouldn’t care about… who gives a f— about CUDA.” Once software abstracts the idiosyncrasies away, “the providers will compete on specs and not on fake moats” — though he concedes the politics: “if you piss off Jensen, you might need to kiss the ring to get back in line.”

5. Training is research, inference is production — and inference goes to 95%

The two obey “fundamentally different tectonic forces”: training wants more of everything and iteration speed — “like changing the wheel of a moving car” — while inference wants less: fewer headaches, and no interconnect if you can possibly avoid it. His analogy: one painting versus a million paintings — stroke speed versus a reliable process.
The number that frames the episode: in five years, “95% inference, 5% training.”
Today’s inference is “a lot of duct tape,” but it is not simply training code running only the forward pass. The biggest unclaimed win is autoscaling: people pay for 1,000 GPUs 24 hours a day “even if there’s nobody on production — this is crazy,” when provisioning to demand is worth “5, sometimes 10x.” Over-buying reserved compute (because on-demand rips you off) creates a fake scarcity on top of the real one.

6. Seven times better isn’t enough — the buy-in has to go to zero

Every chipmaker — AMD, Google, Tenstorrent — has the same GTM problem: switching means abandoning a six-year Nvidia amortization plan and maintaining two stacks. His hard-won metric: “being seven times better, in whatever metric you want, is not enough to get people to switch. People will choose nothing over something. I have stories.” TPUs are the proof: “they’re available, they’re great, nobody cares” — because the buy-in is six months of engineering.
The counter-example is top-down: Microsoft bought AMD’s supply and runs ChatGPT on it, which “puts OpenAI in the green” on inference. The math: 8 H100s fit only two 70B models, and doubling GPUs buys ~10% more inference throughput — “the dirty secret nobody talks about” — while AMD’s memory fits one model per GPU, roughly 4x the efficiency on chips 30% cheaper. “It’s a no-brainer” — but only if you start from the product and permeate down; rent them out bottom-up and “people will not rent them, simple.”
ZML’s thesis is the general solution: make the buy-in zero — “you just say, hey, now it’s AMD — boom, it runs.” Then “if you are 30% better, I’ll switch to you,” and incremental gains finally become sellable.
Pressed to pick a stock: “I would go with Nvidia still… because the supply. But if we play our cards right, hopefully I will come back and tell you to buy AMD as much as you can — or Tenstorrent, if they go public.”

7. SRAM economics: Gro and Cerebras are magnificent and unaffordable

There’s no magic in single-stream speed — it’s SRAM, memory on the die. But SRAM eats chip surface and kills yield: Gro carries 230MB per chip against a 70B model in BF16 at 140GB — “you do the math.” Cerebras’ wafer-scale engine holds 44GB, water-cooled with copper needles touching the chip: “crazy stuff… very impressive technology, mind you, but very, very expensive.” He suspects the Mistral and Perplexity speed showcases “were done at a loss.”
The names he sees going in the direction of breaking the price barrier: Etched and likely VSORA — “if you can deliver this at a price comparable to GPUs, you’ve won.”
On Nvidia’s HBM position, his “Nutella situation”: Nutella owns 80% of the hazelnut market — “yes, you can do a competitor, but who will you buy the nuts from?” Note the nuance: pure SRAM is “a dead end in terms of scaling” — the future needs HBM and more SRAM in better-dedicated architectures.
The frontier after that: compute-in-memory — bring the processor to the data instead of shuttling it. Two companies he names: likely Rain AI (“likely Sam Altman is one of the investors — no surprise”) and Fractile in the UK. “Maybe not this year, but it’s coming.”

8. Products, data, compute — “this is the sleeping giant”

His mental model of who wins AI: a triangle of products, data, and compute. “Who has all three? Google… Android, Google Docs, whatever — they have everything, they can sprinkle everywhere. This is the sleeping giant in my mind” — with the dry caveat: “if they’re not busy doing a reorg.” Amazon has AWS but “they don’t have actual products.”
The shallow take he rejects: that OpenAI kills the search golden goose. “OpenAI is amazing, but it’s not their compute” — it’s Microsoft’s, bought at “outrageous margins.” “If you don’t own your compute, you’re starting with something at your ankle.” Data-center builders buying tens of thousands of GPUs told him the only concession they get is supply — no discounts.
On who actually threatens OpenAI: “I’m not sure who is a threat to OpenAI at the moment.” His mother knows ChatGPT; ask her about Gemini — “what is it?” End state, borrowing Gary at Fluidstack’s analogy: model providers become car makers — no winner-take-all, “everybody will have their own.”

9. Brute force meets its counters: DeepSeek, and Stargate as “an American car”

Scaling today is “a Brute Force approach… a very American approach — more and more and more,” but physics pushes back: the xAI cluster “is not 100,000 GPUs, it is four times 25,000” because interconnect binds it — “at some point you’re fighting physics,” like approaching the speed of light. The two counters he credits: DeepSeek’s (“there’s a lot of waste… virtually you increase your compute capacity because you’re more efficient”) and likely Yann LeCun’s (“this is not scaling — we need to look the problem in the face”).
On DeepSeek itself: “Constraint is the mother of innovation… if you can buy more, why would you give a damn?” China is “a bit late” in terms of ASICs — around A100-level — but it’s exercising in water: “they are bound to do better… it’s short-term to think that way.” The same lens colors his Mistral defense: “it’s easy to spread FUD… what I know is I hope they don’t have too much money.”
The host cites hyperscaler capex of Meta at $60–65B and Microsoft at $80B; Morin says that spending chases training, where Nvidia purchases are incremental — 1,000 Nvidia GPUs plus 1,000 more make 2,000, but 1,000 Nvidia plus 1,000 AMD make “twice a thousand.”
Stargate’s $500B: “My first impression was that I don’t buy it… American style — you start with the claim and figure it out later.” Even if real, it’s more vertical scaling: “an American car of AI — it’s big, it consumes a lot of gas, but ultimately it’s not a good car.” The two limiting factors are “talent and energy, that’s it” — and at 90% margins, “if we work on that margin we can shrink that number.”

10. The model is going away — backends, world models, and better data

His opening frame for where models head: closed-source “models” are already constellations of backends — ask for an image and you’re silently routed to a diffusion model. “Models in the sense of getting weights and running them is ultimately going away in favor of full-blown backends.”
He’s “very bullish” on the likely JEPA world-model approach associated with likely Yann LeCun: “he explained to me how it worked and I was blown away.” Language is “a very narrow window into the world… we describe it arguably with some loss”; energy-minimization thinking matches how he grabs an AirPods case without narrating the motion. Alongside: non-Transformer architectures that “might completely obsolete the Transformers,” and diffusion LLMs that think “in patches,” like humans.
On synthetic data he stays split, exactly as hedged: re-injecting model output “feels wrong… it will deteriorate,” yet AlphaGo’s leap came from synthetic games, and code is one example because “you just run the code” and harvest ground truth — the Poolside thesis. Distillation is “fair game” — a model prompted to generate a Star Wars frame produced the actual screenshot, so “you took it from the beginning too… I take my cup and enjoy it very much, that movie, every single day.” Most mind-blowing: distilled small models sometimes beat their teachers — better “at the task at which you were measuring it.”
Direction of travel is smaller: “a lot of times it’s actually wasteful to run big models,” and people at DeepMind “don’t even fine-tune anymore” — they dump data into huge context windows at runtime. Morin says ZML is working on “attention-level search” as a RAG frontier; Harry calls it an exclusive. That extends RAG’s “clever trick” of prepending semantically-near chunks.

Vol. 52: A 74-Page PPT Breakdown of DeepSeek and AI Agents DeepSeek, Kimi & MiniMax Attention Mechanisms, Paper by Paper