Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN
The AI infrastructure gold rush is here, but scaling it requires navigating colliding forces: billion-dollar GPU contracts, shrinking token economics, specialized vertical models versus general-purpose platforms, and the race to bring compute to the source of power—whether that's West Texas wind farms or modular nuclear reactors. Four CEOs at the frontier of the AI economy open up about their distinct visions. CoreWeave's Michael Intrator reveals how a hedge fund pivoted from crypto mining into building supercomputers for hyperscalers, financing infrastructure through «boxes» that pay for themselves in 2.5 years. Aravind Srinivas of Perplexity makes the case that the orchestrator who can route intelligently across models—without favoring one—holds the real strategic advantage in an era of model specialization. Arthur Mensch of Mistral argues that open-source verticalized models, not closed general-purpose giants, will unlock proprietary data in aerospace, finance, and heavy manufacturing. And Daniel Roberts of IREN explains how the real constraint isn't GPUs anymore—it's time to compute, transmission capacity, and finding enough electricians willing to move to remote Texas sites powered by 100% renewable energy.
Pontos-chave
CoreWeave finances GPU infrastructure through self-amortizing «boxes»—contracts where Microsoft or other hyperscalers pay off principal, interest, and data center costs in 2.5 years of a 5-year deal, enabling rapid scale without traditional equity dilution.
Perplexity's competitive moat is multi-model orchestration: by remaining neutral and routing queries to the best specialist model (GPT, Claude, Gemini, Qwen, DeepSeek), it avoids the liability of betting on a single architecture and can deliver superior accuracy across tasks.
GPU obsolescence fears are «nonsense»—CoreWeave's average contract is 5 years, Ampere A100 prices have appreciated year-over-year, and the installed base continues to find profitable inference and experimental workloads long after frontier training moves to newer chips.
Mistral and enterprise customers are discovering that closed-source models can't leverage decades of proprietary IP; open models enable post-training customization, on-premises deployment, and deep integration with physical systems in manufacturing, finance, and science.
IREN built 4.5 GW of capacity—nearly the Bay Area's annual power consumption—by locating data centers at the source of stranded renewable energy in West Texas, where 45–50 GW of wind and solar exceeds the 12 GW transmission line to population centers, creating an arbitrage opportunity between cheap electrons and high-value compute.
Em resumo
The AI infrastructure race is no longer about who has the most GPUs—it's about who can finance them sustainably, route workloads intelligently across specialized models, and build data centers where cheap renewable power actually exists, all while training a new generation of trades workers to make it happen on the ground.
From Crypto Mining to AI Infrastructure: CoreWeave's Pivot
CoreWeave bootstrapped a neocloud by mining Ethereum, then donated A100s to open-source AI researchers.
Michael Intrator didn't set out to build an AI cloud—he ran an algorithmic hedge fund focused on natural gas. In 2017, he and his team got interested in cryptocurrency and began mining Ethereum with GPUs, viewing compute as an option on multiple use cases. After weathering two crypto winters, they moved up the complexity stack: CGI rendering, batch computing for medical research, and eventually neural networks. In 2020–21, CoreWeave bought A100s and donated them to Luther AI, an open-source project, reasoning that researchers couldn't complain about service-level agreements if the GPUs were free. The feedback loop was immediate: «We need more of this, you got to work on this.» When those volunteers returned to their day jobs at research labs and hyperscalers, they demanded the same infrastructure. That launched CoreWeave's commercial business.
The company understood early that computing decommoditizes at scale—anyone can run a GPU, but can you run a cluster large enough to train a world-changing model? CoreWeave focused on delivering supercomputers purpose-built for AI, living «above the Nvidia GPUs but below the models.» Their first large commercial customer was Inflection; soon they were serving OpenAI, hyperscalers, and foundation-model builders. The thesis: as scaling laws drive larger training runs, only a few players can deliver infrastructure that doesn't bottleneck innovation. Today, CoreWeave is the tip of the spear bringing new Nvidia architectures—H100, H200, GB200, GB300—into commercial production at scale, each generation moving from training to experiments to long-tail inference over a multi-year lifespan.
The «Box» Model: How CoreWeave Finances Billions in GPUs
CoreWeave securitizes client contracts, data-center leases, and GPUs into discrete cash-flow vehicles that self-liquidate.
Sign a 5-year contract CoreWeave negotiates a contract with a hyperscaler or foundation-model customer (e.g., Microsoft) for compute over five years.
Create the «box» The contract, GPU purchase orders from Nvidia, and data-center lease are placed into a special-purpose vehicle (SPV) that governs all cash flow.
Build and deliver compute CoreWeave constructs the infrastructure and begins delivering compute. Microsoft pays the box directly, not CoreWeave.
Waterfall payments Incoming revenue pays the data center, power bill, interest, and principal in sequence; excess flows back to CoreWeave equity.
Self-liquidation in 2.5 years Within 2.5 years of a 5-year deal, all principal and interest is repaid. The remaining 2.5 years generate pure equity returns.
Scale and lower cost of capital Each successful box builds lender confidence. CoreWeave has dropped its cost of capital by 600 basis points in two years, approaching hyperscaler borrowing rates.
GPU Depreciation Debate: «Nonsense» or Real Risk?
CoreWeave's Intrator calls obsolescence fears trader-driven FUD; A100 prices appreciated year-over-year.
“The facts on the ground is they're buying it for 5 years. If people are willing to pay me for it, it still has value. The A100s, the Ampers this year, the price has appreciated through the year.”
Perplexity's Three Strategic Layers
Why Perplexity Believes Multi-Model Orchestration Wins
Models are specializing, not commoditizing—neutrality lets Perplexity take the best of each.
Aravind Srinivas argues that Perplexity's strategic advantage is being «Switzerland»—no allegiance to a single model vendor. As Anthropic CEO Dario Amodei observed, models have begun to specialize: even within coding, Claude Code and Codex have distinct strengths. Perplexity's iOS engineers prefer Codex; backend engineers prefer Claude Code. Across other domains—visual synthesis, audio, multimodal reasoning—different models excel. The company that can intelligently route workloads and orchestrate the «best musician» for each task delivers superior value without the liability of betting on one architecture.
Srinivas introduced «Model Council,» a feature that runs the same prompt across five models, then highlights where they agree, disagree, and where nuances lie—addressing the problem Jensen Huang described of manually querying five AIs and applying human judgment. Perplexity abstracts that complexity. For enterprise customers on the $400/month Enterprise Max tier, every dollar of revenue has positive gross margins because the company isn't just reselling tokens—it's selling orchestration, routing efficiency, and recurring subscriptions. Unlike wrapper companies that subsidize usage, Perplexity's multi-model harness and retrieval-augmented generation (RAG) architecture reduce context-window bloat and token costs, enabling profitability at the unit level even as the company scales.
Perplexity's Enterprise Push: Computer as a Slack Bot
Thousands of companies use Perplexity Computer; internal employees talk more to it than colleagues.
Perplexity's Enterprise Push: Computer as a Slack Bot
Perplexity's fastest-growing revenue stream is enterprise. Computer exists as a Slackbot on the Enterprise plan, and Srinivas reports that «people are talking more to Computer on Slack than to other people.» The company has saved over $100 million in labor costs for Enterprise Max customers. One board memo, one partnership deck, and one press briefing prep were all one-shotted by Computer—tasks that previously required design teams, comms staff, and hours of human coordination.
Mistral's Thesis: Open Models Unlock Proprietary Data
Closed models can't integrate decades of manufacturing or financial IP; open models can.
Mistral's Forward-Deployment Model
Mistral sends PhD scientists to customer sites, trains models on-premises, ensures zero data leakage.
Deploy portable platform Mistral's training tools and data-processing pipelines are installed on the customer's infrastructure, ensuring no data flows back to Mistral.
Send forward-deployment engineers PhD-level scientists work alongside vertical experts (e.g., image-scanning engineers at ASML) to understand workflows and defect-detection logic.
Ingest proprietary data Customer data—often decades of manufacturing signals, medical images, or financial transactions—is used to fine-tune open-source models.
Train and hand off Models are retrained on-site. Once complete, the customer can operate independently, retraining and iterating without Mistral's ongoing involvement.
IREN: Building Data Centers Where the Wind Blows
IREN co-locates compute with stranded renewable energy, avoiding transmission bottlenecks and achieving 100% renewables.
Daniel Roberts and his brother started IREN eight years ago in Sydney, betting that the digital world's exponential growth would strain real-world infrastructure. Their thesis: build large-scale data centers, bootstrap with Bitcoin mining cash flow, then layer in higher-value use cases as they emerge. Today, IREN is swapping out all Bitcoin miners for AI chips, and the company has 4.5 GW of capacity under development—nearly the annual power consumption of the entire San Francisco Bay Area.
The breakthrough insight: West Texas has 45–50 GW of wind and solar generation, but only a 12 GW transmission line to export power to Dallas and Houston. That creates stranded renewable energy. IREN builds data centers at the source, monetizing excess wind and solar into compute and exporting it «at the speed of light as tokens.» The utility underwrites variability and guarantees 24/7 reliable power; IREN pays for 100% renewables without needing to manage batteries or backup generation. Four years ago, when IREN proposed a 750 MW site in the middle of the Texas desert, the traditional data-center industry thought they were crazy. Today, the company has signed a $9.7 billion contract with Microsoft—representing just 5% of its total capacity.
Key Infrastructure Metrics
CoreWeave, Perplexity, and IREN shared hard numbers on contracts, capacity, and costs.
The Real Constraint: Time to Compute
GPUs are available—but building data centers in remote Texas requires thousands of tradespeople.
IREN's Daniel Roberts argues that the industry narrative has shifted: the constraint is no longer GPU supply, it's «time to compute.» For IREN, which secured land and power eight years ago, the bottleneck is physical: breaking ground, pouring foundations, installing water-cooling systems, running fiber, and hiring electricians, HVAC technicians, and construction crews. These are hard manual-labor jobs in remote locations, often requiring workers to live in temporary housing for multi-month tours of duty.
IREN's policy is to hire locally first, expanding the radius in 20-mile increments as needed. They target areas where heavy electrical infrastructure already exists—typically where old manufacturing has shut down—and retrain local workforces. The company is partnering with trade schools and universities to build a new generation of skilled labor. Starting salaries for electricians and construction workers on IREN sites are in the $150K range, making trade careers financially competitive with white-collar knowledge work. Roberts sees this as a reversal of the decades-long offshoring of manufacturing: AI infrastructure is bringing high-paying blue-collar jobs back to rural America, powered by the same wind and solar farms that replaced coal and gas in the energy transition.
Jevons's Paradox and the Infinite Demand for Compute
As inference gets faster and cheaper, humans will generate exponentially more queries—not fewer.
Jevons's Paradox and the Infinite Demand for Compute
Roberts invoked Jevons's paradox: if generating an image in ChatGPT today takes two minutes, users carefully craft prompts and generate sparingly. If a 10× increase in compute drops that to 5–10 seconds, will users generate more or fewer images? Many more. The same applies to autonomous vehicles, real-time translation, and agentic workflows. Faster, cheaper inference doesn't reduce demand—it unlocks entirely new use cases and consumption patterns, creating a positive feedback loop that drives exponentially higher compute requirements.
Títulos mencionados
Pessoas
Glossário
Aviso: Este é um resumo gerado por IA de um vídeo do YouTube para fins educacionais e de referência. Não constitui aconselhamento de investimento, financeiro ou jurídico. Verifique sempre as informações com as fontes originais antes de tomar decisões. O TubeReads não é afiliado ao criador do conteúdo.