Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN
Summary
CoreWeave’s central claim is that AI compute decommoditizes at cluster scale, while contractual cash flow lasts far longer than the market’s feared chip cycle. Michael Intrator called 16-month-or-so GPU obsolescence “nonsense”: the average contract is five years, CoreWeave depreciates equipment over six, expects useful life beyond that, and says A100 pricing appreciated during the year. Hardware becomes obsolete when data-center power can earn a higher margin elsewhere.
CoreWeave’s growth engine is structured finance, not uncontracted GPU speculation. Each discrete “box” contains a customer contract, GPUs and a data-center agreement; customer payments cover facilities, power, interest and principal before residual cash returns to CoreWeave. That structure supported $35 billion of financing in 18 months, pays for everything within 2½ years of a five-year deal and helped lower its cost of capital by 600 basis points.
Infrastructure demand remains above physical supply, but the bottleneck now spans memory, power, optics, networking and construction—not merely NVIDIA allocations. Intrator said four years of demand have “overwhelmed the global capacity of the world,” while IREN’s Daniel Roberts said there are “no idle GPUs in the world sitting in a data center.” Long-term contracts protect CoreWeave against an air pocket; IREN’s position is 4.5 GW of power and an eight-year head start tying up land and power.
Perplexity is betting that value migrates from any single frontier model to the neutral conductor that selects and coordinates them. Aravind Srinivas called the company “Switzerland”: GPT, Gemini, Claude, Kimi, Nemotron and Qwen can specialize while Perplexity auto-routes work, or its Model Council explains where multiple models agree and disagree. The company serves several tens of millions of consumers, says enterprise revenue is growing faster, and reports positive gross margins on “every single penny” of revenue despite not yet being profitable overall.
Perplexity’s road map turns AI from an answer box into the computer and eventually the operating system. Personal Computer will use a Mac mini as a private local runtime while delegating permitted frontier-model calls or long tasks to a user-isolated server; Srinivas described it as “Open Claw for dummies.” The interface starts with objectives rather than programs, while Linux, files, models, connectors and sandboxes sit beneath the orchestration layer.
The near-term disruption is bespoke software and autonomous back-office work, not a single all-purpose chatbot. Perplexity Computer already produced the company’s board memo, partnership deck and press briefing; Srinivas’s longer-term target is a small business whose AI runs ads, support, billing and feature development while its owner is “sipping wine in Napa.” He stressed that this future “is not there yet,” and conceded temporary job displacement even as he argued it could restore agency to people who dislike conventional jobs.
Mistral’s enterprise thesis combines open models, human expert signal and strict execution controls. Arthur Mensch said synthetic data can warm up and compress a smaller model, “but eventually you do need to have human signal”; Mistral therefore deploys portable training tools and PhD-level engineers inside customer infrastructure so data need not flow back. For production agents, OpenClaw-style autonomy is insufficient without deterministic gates, sandboxes, role-based controls and a “context engine” that prevents compensation or other restricted data from leaking across an organization.
IREN’s legacy power portfolio has become its scarce AI asset, with a $9.7 billion Microsoft contract consuming only 5% of capacity. Its 750 MW Texas flagship sits inside a 4.5 GW portfolio, while its operations have used 100% renewable energy since inception—British Columbia hydro and West Texas wind and solar. Roberts’s demand model is Jevons paradox: if 10 times more compute cuts image generation from minutes to 5-10 seconds, users will generate many more images, not bank the efficiency savings.
Deep dive
Not yet available upstream; scheduled sync will retry.