November 21, 2026
Heard in AI

Investor Lonis Hamaili argues AI's value will flow to its factories

Lonis Hamaili, whose Cambrian fund invests in AI's physical supply chain, argues that memory, wafer and fab makers will capture more value than AI labs and are underpriced. He concedes the bet pays off only if AI demand really arrives.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on The Cognitive Revolution, episode published October 24, 2026

Lonis Hamaili sorts the AI hardware business into two camps. One designs chips: NVIDIA, Broadcom, Arm and Marvell. The other makes the physical things those designs depend on, such as silicon wafers, memory and the huge factories, called fabs, where chips are produced. Hamaili is managing partner of Cambrian, a fund that invests in the physical supply chain behind AI, and he is "way more bullish on the atoms over the bits."

He made the case on The Cognitive Revolution, in an episode published October 24, 2026, with host Nathan Labenz and co-host Prakash Narayanan. Labenz told listeners he has committed to invest in Cambrian.

Designers versus makers

Hamaili called the big chip designers great, well-run businesses. But they own no factories and have no physical manufacturing know-how of their own, he said. They send their designs to Korea, Japan and Taiwan, where manufacturers have built up decades of experience. That leaves the designers with a much weaker moat, investors' term for a company's protection against competitors.

He said the AI labs already compete directly with designers like NVIDIA and Broadcom. Wafer makers are another matter. "There will never be a situation where the labs will compete with a silicon wafer maker," Hamaili said. "And we will always be a customer."

Cambrian's published thesis sets out the framework behind this view. It maps fifteen layers of AI infrastructure, from commodities and electricity through fabrication equipment, memory and packaging to cloud providers. Companies are judged on two separate questions: how exposed they are to AI demand, and how easily someone could replace them.

Why he says the market hasn't priced it in

Labenz started with the obvious objection. The AI boom is "all staring us right in the face," he said, so why wouldn't everyone already know all this?

Hamaili answered with valuations. By his figures, semiconductor stocks trade at about 16 times earnings, while the S&P 500 averages about 22 and Walmart almost 40. A price-to-earnings multiple roughly tells you how many years of current profit investors are willing to pay for. Hamaili read the low multiple as a judgment that chip earnings will not grow and are less durable than the rest of the market's. The companies Cambrian likes best, in silicon wafers, memory, capacitors and foundries (contract chip factories), trade at three to five times sales, he said.

He offered two reasons for the gap. The first is that most investors are not "AI-pilled," meaning convinced that AI will transform the economy. "Even in the Bay, some people are not AI-pilled. But if you go to the East Coast, you know, almost nobody is AI-pilled," he said. The second is a Cambrian view that private markets price AI more efficiently than public ones. In private deals, former operators and technical people bid startups up to 50 or 100 times sales, sometimes for good reasons. On Wall Street, he said, technical believers control only a tiny share of the money being traded.

Narayanan holds a related view. In a clip Labenz introduced as his co-host on "what the market keeps getting wrong," Narayanan said he expects markets to underprice exponential growth for the next 20 years. People cannot understand it or get emotionally comfortable with it, he said, so markets keep swinging through a "bubble moment, collapse moment" like the one of the previous three months.

His own best case against the bet

Labenz asked Hamaili to steelman the other side, meaning to make the strongest case that these companies are fairly priced. Hamaili conceded that many of them are not great businesses in a normal world. They have high fixed costs, sometimes heavy debt, and they lose money when their factories run below a certain utilization rate.

His illustration showed why the payoff depends so heavily on AI. Take a company with 10% gross margins, "a terrible business by Wall Street standards." If it can raise prices by 10%, its earnings double. If it can double prices, he said, earnings rise tenfold. That only happens "in this very specific AI world," Hamaili said. In normal times these companies are good at matching supply to demand, so prices almost never jump like that.

Memory supply he says can't triple

Narayanan, Labenz said, had raised the market's main worry: an oversupply of memory by 2028 or 2029. Hamaili said memory needs few things to go right. It already trades at five or six times earnings and should eventually be re-rated to at least 10 or 12, he said, even without further price increases, provided demand keeps outpacing supply.

On supply, he said almost no expert thinks the industry can grow DRAM output by more than 2.5 times in bits by the end of the decade. DRAM is the fast working memory that AI systems depend on. No expert with a track record, he said, has claimed supply can triple by then. A naive bear case therefore has to assume that demand will not triple either, "which, of course, everyone here would assume we will."

The wild card is compression: how many bits of memory a model needs for each token, the small chunk of text it processes. Cambrian has tracked this from GPT-3 to DeepSeek-V4. Over that span, Hamaili said, memory bits per token fell about 500-fold, yet the price of a DRAM bit rose six- or sevenfold. DeepSeek-V4's documentation describes one way such savings are made: the model keeps nearby text in full detail while holding compressed summaries of more distant text, which shrinks the history it must keep in memory.

Asked what would make memory prices fall, Hamaili named two scenarios. Compression might improve much faster than other economically useful capabilities, perhaps once AI becomes a superhuman coder. Then cheaper memory per token would not be offset by much greater use, the effect known as the Jevons paradox. He called that relatively unlikely. The other scenario is that energy becomes a bigger bottleneck than chips. Buyers would then have less reason to bid up the last wafer or memory bit if they must wait a quarter for power anyway. Parts of 2026 looked like that, he said.

Why he thinks power is softer than silicon

Hamaili expects that period to end. Data-center builders have learned to work "behind the meter," building their own natural-gas power plants instead of waiting for a grid connection. The large turbines that turn gas into electricity are sold out until 2030 at GE Vernova and Siemens, by his account, but builders improvise with truck engines and parts from airplanes. Energy has spare capacity in odd places to draw on. Chipmaking does not. "There's no latent supply of wafers or of memory bits," Hamaili said. "There's no fat, essentially."

Can superintelligence route around the bottleneck?

The hardest pushback came next. If AI becomes vastly more intelligent, the question went, couldn't a relatively inexperienced company in a second-tier country suddenly gain the ability to make semiconductor equipment, as people predict for biology?

Hamaili disagreed, and his reason was training data. AI models can learn from a large body of biological data. For making a silicon wafer or DRAM, he said, there is "not really much good training data at all." Most of that knowledge is tacit and "sits in every engineer's head." There are no millions of pages of manuals online and no reinforcement-learning environments, the simulated practice tasks in which models are rewarded for getting things right.

The discussion pressed further: the labs are buying data from semiconductor companies and building their own training environments, and OpenAI has a deal with "I think, Synopsys or Cadence." That deal exists. On September 30, 2026, OpenAI and Synopsys announced a multiyear agreement to build GPT-Synopsys, a model that operates chip-design software, with revenue shared between them.

Hamaili said that fit his distinction. Designing chips is a core priority for every lab, he said, and one of the main ways to close the loop of recursive self-improvement, in which AI improves the tools used to build better AI. Designing and building a fab of two million square feet from scratch is different. That knowledge sits with people like a Japanese engineer who has worked at the same company for three decades, and it is hard to extract in structured form. Superintelligence will eventually work out how, he said, "for sure." What matters is the timing and the order in which things happen.

Huawei was his example. He said the Chinese company can design chips of roughly the same quality as AMD and NVIDIA. The difference is that NVIDIA has TSMC's factories and Samsung's high-bandwidth memory, while Huawei is limited to the less advanced manufacturing processes available in China and older memory generations, HBM2 and HBM3.

The last challenge was that intelligence is a way of finding routes around obstacles, so its first big job would be to clear the supply chain's bottlenecks. Hamaili agreed that "will certainly happen." But an AI smart enough to solve some of the hardest engineering problems in the world could also automate many of the "email jobs and Excel jobs" that cost trillions of dollars a year, he said. That would unleash trillions of dollars of value "that the supply chain could capture."

Does Elon Musk take anyone with him?

After the interview, Labenz and Narayanan kept talking. Narayanan had been arguing that Elon Musk treats the industry's fixed constraints as illusions. Asked whether he rejected the whole bottleneck thesis, he said it is "not actually that true." A dedicated "physics guy" can get rid of those bottlenecks, he said, and Musk has capital he is willing to gamble.

Incumbent suppliers rarely gamble, Narayanan argued. He doubted that a small Japanese manufacturer worth $30 million would spend $10 million on an entirely new approach. Senior management at ASML, the Dutch company that makes the lithography machines used to print the most advanced chips, would struggle to redesign around a paradigm its customers don't want, he said. Those machines are hard to reinvent: ASML's explainer describes a light source that fires laser pulses at droplets of molten tin, 50,000 times a second, to produce the extreme ultraviolet light its newest machines use. "Elon will risk like a hundred bill," Narayanan said. "He can absorb the losses. ASML cannot."

Labenz asked whether Musk, if he escapes gravity "literally and figuratively," takes anyone with him, or whether everyone else stays gravity-bound and pays the market price for scarce components.

Narayanan pointed to rocket launches. By his figures, Musk holds 90% of the launch market even before Starship and will hold 99.9% of capacity once Starship flies next year. He said the remaining rivals are largely state-backed or funded by the stock market or Jeff Bezos, and are not really profitable. Musk could end up with 99% of the semiconductor market the same way, he said. He does not expect it, because he believes in recursive self-improvement and that Anthropic and OpenAI will reach superintelligence and then be able to buy or create much of this supply chain themselves. "If not for the RSI thesis," Narayanan said, "I think Elon will win."

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04

Connected ideas and articles

From the conversation

Podcast episodes

The Cognitive Revolution

AI:AM Highlights: Clean Neurons, Black October, the GPU Credit Crunch, and Buying the Supply Chain

Episode published This article draws on 1:40–3:31 and 51:58–1:07:24 (approximate times)

Article history

Updates to this article

Tags