Michael Intrator, Aravind Srinivas, Arthur Mensch: Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN

The episode is recorded at NVIDIA's annual GTC conference, where the host sets out to interview four AI CEOs. The first conversation is with CoreWeave, introduced as a company building massive infrastructure for hyperscalers and described by the interviewer as "the original hyperscaler" because it moved into GPUs so early. The opening question is how the company got there so early and how it built what the interviewer calls, with some hesitation, the first "Neo Cloud".

The answer begins by rejecting the premise that CoreWeave started as a Neo Cloud. The founding story runs through an algorithmic hedge fund focused on natural gas: once the algorithms were built, the work was mostly monitoring and testing, which left downtime. The team, self-described as "pretty nerdy", got interested in crypto and especially its security layer.

Bitcoin mining was ruled out. The reasoning given is that whoever engineered the ASIC would probably run it better than they could. GPUs were more attractive because they could mine Ethereum but could also be turned to other work. From the start, compute was treated as an option that could be redeployed across use cases. The company began in 2017, spent roughly its first three years mining crypto, and went through a couple of crypto winters.

The hedge-fund background is credited with how the company survived those winters. The account points to experience in risk management, capital, risk exposure and allocation, and says the company was careful about these from the start. Because crypto was volatile, the company immediately looked for other uses for its compute. The interviewer adds that crypto was "a question mark" then and that the only other users of this hardware were quants and medical researchers.

The product progression is laid out step by step. After crypto came CGI rendering for people animating and rendering images, then batch computing and medical research, each step more complex in how GPUs were used. Around 2020 to 2021, the company tried to learn how to run neural networks on GPUs. It did not know how, so it bought a batch of A100s and donated them to a group working on Luther AI, an open-source project.

Donating the hardware was described as a deliberate hedge: users who receive free compute "can't really get pissed at us if we're not very good at it." The interviewer jokes that they "can't complain about the SLA." The volunteers kept asking for more and pointing out what needed work, and that feedback taught the company what it takes to run parallelized computing at scale.

I kind of feel like buying those initial GPUs was the tuition we paid to learn how to run this business.

There was also an unplanned commercial payoff. The volunteers were like-minded scientists who went back to their day jobs and wanted the same infrastructure there, because it was "built the right way" for researchers. That demand is described as what launched the business.

Asked about the next card in the "poker game", the account goes back to 2020 to 2021, before the ChatGPT moment. The company concluded then that scaling laws would drive the field and that large-scale compute behaves differently from commodity hardware.

Anybody can run a GPU, but can you run a cluster that's large enough to train a model that can change the world? And that's a different question.

The strategic shift was to treat the work as a business that needed access to capital, so it could deliver to larger clients and the most sophisticated buyers of compute, rather than as an engineering project. CoreWeave places itself above the NVIDIA GPUs and below the models. Its layer covers software, integration, operations and observability, all purpose-built for one use case. Asked about web servers, the speaker credits AWS with "a brilliant solution to solve a problem" and presents CoreWeave as a solution to a new problem.

On customers, the first language-model user is given as Luther. The first large commercial customer is given as Inflection, working with Mustafa. From there the business diversified into hyperscalers, OpenAI and other foundation-model builders, always aiming to build "supercomputers that can change the world."

The conversation then moves from training to deployment. The account says AI has gone from research to productization and is moving from the fringe of organizations into their core work. As evidence it cites the volume of inference running through CoreWeave's infrastructure, which it calls "just massive." No figures are given.

I always think of inference as the monetization of the investment in artificial intelligence.

Inference is defined in plain terms: asking a model a question and getting an answer, or asking it to go and do something. It is presented as the point where value moves beyond the model into the real world, and as the metric watched to gauge the health of the business.

CoreWeave describes itself as "the tip of the spear" in bringing new NVIDIA architectures into commercial production at scale. It claims to have been first at scale with H100s, H200s and GB200s, with GB300s now arriving. The described lifecycle is that customers train on the newest GPUs, then move them to other experiments, and eventually to inference, where they stay in use "for a very, very long time."

The interviewer raises the debate over how long this hardware lasts. The question names the company, Microsoft, and Michael Burry, whom the interviewer characterizes as saying "the sky's falling." The interviewer's counterpoint is that the industry does not throw the hardware away: "The street finds its own use for technology."

So my take on the GPU depreciation debate is that it's nonsense, right?

The rebuttal has two parts. The first is about motive: the speaker says the debate is being pushed by traders with short positions trying to talk the stock down. The second is about contracts: clients buy compute for five or six years, and the average contract is five years. Claims that the hardware is obsolete in sixteen months are said not to match "the facts on the ground." The underlying test is "if people are willing to pay me for it, it still has value."

Concrete claims offered in support: - CoreWeave uses a six-year depreciation schedule and believes the GPUs will last beyond six years. Six years is presented as a fair and reasonable choice given how fast the technology cycle is moving. - The price of A100s (Amperes) has appreciated through this year. - The explanation given is that new installed capacity attracts new companies with new use cases and different model sizes, including some that were shut out of H100s. - Six years is described as the industry standard, with Amazon named as the one exception. The speaker adds, "I'm not making it up."

The interviewer compares this to second-hand phones: an iPhone 12 or Pixel 7 sold for $50 in South America or Africa still has useful life left. The speaker adds that new companies, and existing companies adding new models to their workflows, keep buying any Ampere GPUs available. Both sides call the idea that GPUs stop being commercially viable after sixteen months, eighteen months or two years "farcical." The interviewer suggests that Moore's Law and the industry's growth push big companies to demand the newest products. In that view, the opportunity has grown larger without the hardware's lifespan getting shorter.

The speaker also explains why the debate gets attention. The unprecedented scale of capital draws focus to companies building on the most advanced chipsets. Yet even those companies keep a long tail of uses for older hardware: inference, other experiments, and less cutting-edge work that still needs to be done. The interviewer adds rendering and image generation on Nano Banana as examples, and concedes there may come a moment when the ratio of compute to power no longer makes sense.