Four CEOs on the Future of AI: CoreWeave, Perplexity, Mistral, and IREN
In a Nutshell
Four AI CEOs at Nvidia GTC detail their pivot from crypto mining to massive AI infrastructure: CoreWeave deploys bleeding-edge Nvidia GPUs via innovative "box" financing for 5-year contracts with hyperscalers like OpenAI, emphasizing inference monetization and 6-year hardware lifespans amid relentless demand constrained by power, memory, and data centers; Perplexity evolves from search to a hybrid AI "computer" with browser/root access, multi-model orchestration, and enterprise tools enabling one-person $1B businesses; Mistral focuses on open, specialized European models via Forge for secure enterprise customization; IREN swaps Bitcoin for AI data centers with 4.5GW renewable power, partnering Microsoft amid talent shortages and Jevons paradox-driven compute explosion. Key takeaways: AI demand overwhelms global capacity, inference drives value, capitalism's boom-bust cycles (e.g., memory, energy) fuel innovation, and hybrid local/server agents lower barriers for entrepreneurship while demanding robust governance.
These notes were generated by AI and may contain inaccuracies.
Interview with four AI CEOs at Nvidia's annual GTC conference: CoreWeave, Perplexity, Mistral, and IREN. Sponsored by the New York Stock Exchange, a modern marketplace for raising capital and building for the future.
CoreWeave, building massive infrastructure for hyperscalers. Michael Intrater, CEO, describes starting as the original hyperscaler by securing GPUs early. Began running an algorithmic hedge fund focused on natural gas. After algorithms were built, downtime led to interest in crypto. Examined Bitcoin mining but preferred GPUs for Ethereum mining due to versatility for other uses. Started company in 2017, mined crypto for first three years, weathered two crypto winters using hedge fund risk management expertise in capital allocation and exposure.
Scaled company and sought other use cases due to crypto volatility. Progression of products: crypto to CGI rendering for animating and rendering images in movies, then batch computing for medical research and science. Moved up the stack in GPU complexity. In 2020-2021, explored GPUs for neural networks. Bought A100s and donated to Luther AI open-source project volunteers. Learned scale parallelized computing from their feedback. Those researchers returned to day jobs demanding similar infrastructure, launching CoreWeave's business.
Recognized scaling laws would drive demand pre-ChatGPT (2020-2021). Compute decommoditizes at scale: anyone can run a GPU, but running world-changing model clusters is different. Focused on scaling delivery to larger clients, accessing capital for sophisticated consumers. Thinks of business above Nvidia GPUs but below models: software integration, operations, observability for AI-specific cloud. Doesn't do general web servers (AWS excels there), solves new AI compute problem.
First language model customer: Luther AI. First large commercial: Inflection with Mustafa. Diversified to hyperscalers, OpenAI, foundation models. Builds supercomputers to change the world, now shifting from research to productization. Massive inference compute usage shows models being deployed and utilized, not just built.
Inference is the monetization of the investment in artificial intelligence.
Inference: asking model a question for answer or action, driving real-world value.
Tip of the spear deploying new Nvidia architectures at scale: first with H100s, H200s, GB200s, now GB300s. Bleeding-edge GPUs used for training, then experiments, then long-term inference.
GPU depreciation bait is nonsense from short sellers. Success-based company with 5-6 year average contracts; clients buy compute for 5 years. Uses 6-year depreciation, expects GPUs to last beyond. A100 and Ampere prices appreciated this year as new companies/use cases emerge for older hardware (e.g., smaller models, new ventures blocked from H100s). Analogy: traded-in iPhone 12 sells for $50 in South America/Africa with great life left. Lifespan hasn't shortened; opportunity surface area has expanded. Even advanced companies use older GPUs for inference, experiments, rendering. Obsolescence defined by power/data center repurposing for higher margin. Industry standard: 6 years (like Amazon). Energy cost drives replacement.
Affirmation of healthy business with competitors entering. Accesses GPUs like everyone: submits PO and pays. Nvidia allocates in order received, avoiding favorites or auctions. Works with interesting companies fitting long-term contracting/debt structuring requirements.
Pioneered GPU-based loans. Process:
- Secure client contract (e.g., Microsoft).
- Create "box": put client contract, Nvidia GPU purchase, data center contract inside.
- Box governs cash flow waterfall: client pays box (not company), which pays data center/power, then interest/principal, remainder to CoreWeave.
Time-tested vehicle borrowed against client paper/collateral. Raised $35 billion in 18 months. Economics: in 2.5 years of 5-year deal, principal/interest paid off, generating returns. Lenders confident in repayment (rule of lending). Boxes independent, no contagion. Cost of capital dropped 600 basis points over two years toward hyperscaler levels. Militant about maintaining boxes for capital access. Says no to short-term deals (e.g., 1-year) unable to amortize expenses.
Business built on risk management for scale during GPU shortage disequilibrium. Gets large to lower cost of capital, gain info flow from LLMs, high-speed trading, search on next products/needs.
Demand for service relentlessly overwhelms global capacity. Depth steadfast for years. Protected by 5-year contracts with strong-balance-sheet counterparties, even if air pocket from breakthrough/war. Lenders confident (rates declining). If unconstrained with infinite GPUs/energy/data centers:
Constraints aren't just GPUs: electricity, power, shells, memory, storage, networking, optics. Various throttles limit capacity. Memory is a current throttle.
Memory has historically been a cyclical business with waves of demand driving up costs, then collapse. Fabs are capital intensive, leading to overbuilding during turndowns. Currently, AI demand for compute and ancillary GPU services is through the roof. There was an investment cycle needed in 2023 for fab capacity.
Impossible to predict energy needs. Data centers go where energy is, not based on real estate. Chasing energy, some wind.
Capital intensive businesses like fabs and energy lead to boom-bust cycles with overbuilding, like fiber.
Capitalism's beauty: boom-bust cycles weather challenges, clear underbrush, survivors take advantage, sow seeds for future business. Excess fiber became backbone for watching movies, communication, Zoom.
YouTube founders Chad Hurley and partner realized storage costs dropping quickly for free unlimited uploads, bandwidth dropping so no charge for video sharing. Before, viral videos crashed servers due to per-megabit carriage charges.
Business models evolve with Moore's law. Accelerated computing dwarfs Moore's law, leading to more world-changing companies like YouTube.
Susan Wojcicki noted massive uploads per minute, illogical until realizing 3 billion users, 1% or 0.1% uploading makes sense—one in a thousand.
Sarah Friar, CFO of OpenAI, on panel: cost of a million tokens with ChatGPT-3 was $32+, now 9 cents. Capitalism fuels engineering and competition.
Models now recursive: tell model to make itself more efficient, lower token costs. Karpathy's recursive thing inspires civilians to try recursion.
AI lowers barrier to operations. Good ideas can vibe code, create new things. Brings down barriers containing human creativity, enabling new medical research, baseball cards, etc. Valuable kernel is creative ideas to build new things. Exciting: tool for 8 billion minds to overcome insurmountable challenges. Bright new future.
User fell in love in three stages:
- Pick language model (OpenAI, Claude), sidebar like early Yahoo (finance, sports), live personalized summaries for stocks/news.
- Comet browser: instructions like go to LinkedIn, find everyone from company, put in Google Sheet.
- Computer: strong start for repetitive tasks, like co-work from Claude or engineer/developer.
These are company evolution. Perplexity has loyal fan base, making hundreds of millions to billions.
Built for curious people seeking extra edge. Common theme: accuracy to build trust for next questions.
- Perplexity Ask: AI with internet access for accuracy.
- Comet: full browser access for tasks.
- Computer: full computer access, AI becomes the computer.
Orchestra of all AI capabilities. Sub-agents are musicians, models are instruments (hundreds: coding, writing, multimodal, image/video/audio generation). End output matters: work AI gets done.
Lives in browser now. Considered desktop root access, but security/trust issues (hacking, deleting files).
Announced Perplexity Personal Computer: synchronizes server-side trust/reliability with local computer. Use from phone, sync with Mac Mini as local server for private data orchestration. Pings frontier models with permission if needed. Hybrid local/server-side, abstracted—one executable install, no API keys, billing, access management. Steve Jobs end-to-end integration.
Local models like Kimi 2.5 on Mac Studio: 80% capability for free, compelling vs. expensive Claude bills.
Desktop returning to workstation/server status promising. Starts as sub-agent for local data (tax returns, photos, emails, calendar, notes) on local hardware for privacy.
Hybrid orchestration: local for device data, server for cloud data (Google Calendar, Gmail via connectors).
Not dichotomy local vs. server—choice. On phone, workloads run on Mac Studio/Mac Mini/Dell workstation (Dell/Nvidia: giant with 750GB RAM) or server.
$10,000 desktop lowers $500/month Claude bill—incredible savings, privacy, no data training.
Costs drop like refrigerator/internet modem. Homes have sensors for orchestration—dictate to phone to control home.
AI is the operating system. Start with objectives, not instructions. Example: build website tracking All-In podcast transcripts' stock price impact before/after, charted over time.
OS as models, systems, files, connectors: file system, code sandbox, internet, HTML tools. Operate at objective abstraction.
Could become own OS. Perplexity Computer runs on Linux machines server-side. Marc Andreessen: Linux computers right idea, desktop Linux works—stable, customizable, not at Apple's mercy or Microsoft's hacker surface.
Access Linux runtime from phone (iOS/Android), front end irrelevant.
Consumer success, now corporations using Computer. Example: venture firm back office automation only using Perplexity Computer (Slack connector needed, but already out as Slackbot on enterprise plan). Entire Perplexity company talks more to Computer on Slack than people.
Consumer: tens of millions monthly users, like Google/Yahoo trajectory.
Corporate: thousands of companies, fastest growing revenue. Saved enterprise Max customers >$500 million.
Tiers: Enterprise Pro $40/month, Enterprise Max $400/month ($5,000/year). After credits, pay for token usage.
Every revenue has positive gross margins—not just selling tokens. Recurring subscriptions. Efficient via multi-model routing, RAG, orchestration, search—no need to blow context windows.
Company not yet profitable, working toward it.
400-person team, product-driven, understands consumer/business. Declined exit offers (Apple, others).
Competitors: OpenAI ($100B raise), Elon (data centers in space, xAI/Twitter/SpaceX), Google (unlimited resources, strong consumer), Amazon/Gemini, Meta (copies, good at ads).
Advantage: multi-model orchestration like Switzerland—no one horse. Works if GPT, Gemini, Claude, Llama, open source wins. Has Kimi, Neotron, Quen (Alibaba).
No other can orchestrate best of each. Wouldn't admit capex/data centers don't produce best model.
Dario Amodei (Anthropic): models specializing, not commoditizing. Even in coding, Claude Code vs. Codex differ—iOS engineers love Codex, backend love Claude Code.
Orchestra conductor wins uniquely.
Buy tokens wholesale? No, ping models directly. Perplexity provides orchestration/harness.
Auto-routes best model per prompt, user flexibility to pick.
Model Council: same query across multiple models (like Jensen does with 5 AIs). Avoids manual biological trust judgment—like 5 doctors, but still dumb.
The model council feature shows answers from each model, highlights where they agree, disagree, and nuances. It's in the interface.
Perplexity ships at a great cadence. Philosophy: speed is key. Big companies can't move fast or serve customers quickly while maintaining quality, speed, and trust. Apple takes long to ship due to trust concerns. Bureaucratic companies ship slowly, don't maintain products, abandon features after events.
Perplexity's advantages as a small company: AI coding tools speed up shipping. Built computer so non-engineers ship code via Slack bot fixing bugs. Iteration is exponential.
Claude moment: Built a network mind map from 20 Japan contacts via LinkedIn. AI offered outputs: Google Sheet, Notion table, PDF, CSV, or custom CRM. Created a CRM system. Rare experience: agent builds bespoke software (1 in 1,000 or 10,000 users).
Computer examples:
- Made board meeting memo.
- Created partnership pitch deck (one-shot, no design team).
- Wrote press briefing memo (one-shot). Context excels: knows prior journalists, board decks.
Opus 45 enabled orchestration, reasoning, tool calls. Cloud code introduced sandbox/terminal with CLI tools (no GUI needed). Organizes around files, sub-agents, skills, CLIs. Context window no longer an issue; model loads/dumps context as needed. Excels at long orchestration tasks.
Processed every This Week in Startups and All-In episode transcripts. Downloaded all All-In podcasts, extracted public company mentions, histograms, time-series charts, stock price impact, sentiment analysis. Gave timestamps for clips.
Example insight: Google's stock rose after heavy podcast mentions. Public bet: buying Google despite lag, as they're "too precious" and need to release more. Told Sergey to release good stuff.
Research notes mega-prompt: background, circle, competitors, friends, old podcasts (5-10 years ago). Used in interviews like Michael Dell discussing 1990s topics. Replaces $70-80k/year researcher (10x slower, 1/3 quality).
Help businesses run autonomously. Enable one-person $1B companies creating real GDP value. Small businesses: Mac mini or server with Perplexity personal computer. Run ad campaigns (Instagram/Google), SEM/SEO, Stripe integration, Intercom support. Scale to hundreds of thousands or millions revenue while sipping wine in Napa.
Past startups needed office, servers, HR, PR. Now 3-person teams from A16Z Launch, YC raise $0.5-1M, no hires. Handle recruiting: ad, sort/rank candidates, email top 10 questions, interview last 2.
Computer abstracts jobs, needs connectors/workflows. No essay-long prompts; quick/autonomous setup. Turn idea into revenue-generating business.
Enables solo entrepreneurs to $1M revenue without hires. 70-80% positive, 20% worried about fast job loss. Counter: America is about entrepreneurship/exploration. Most hate jobs (Henry Ford factories boxed people). AI enables mini-businesses for needs, high life quality, agency, passion.
Temporary displacement worth it for glorious future. Requires rugged individualism, resilience, resourcefulness. Tools bring out best.
Comet for iOS launched. $300/month initially, now free; Pro $20/month. Integrates with computer on Perplexity app.
Comet uniqueness: controls browser as tool for computer (with Google Workspace). Automates web tasks: tabs, forms, clicks, uploads. Server-side browser execution unique vs. OpenClaw/Claw coworker.
Dream: Android app roots phone for full control. Blockers: sites like Reddit, LinkedIn panic, block bots despite user support.
Happy to work with any company via official APIs in computer. User idea: Pro accounts ($20/month Reddit, $50/month LinkedIn) authenticate tools for limited well-behaved actions (e.g., summarize subreddits like This Week in Startups/All-In for feedback, read-only, named agent). NYT: 100 searches/day on paid sub makes sticky.
Win-win: API access revenue stream. Users choose; applies to any site.
Arthur Mensch, CEO of Mistral AI.
Training next-generation frontier models with Nvidia (like MLMO 18 months ago). Produce best open source models, specialize via Forge for enterprises (engineering, physics, science, languages, governments).
25% business in US, 25% researchers there. Also UK, Singapore. Europe: language focus, more manufacturing. Help lagging European companies leapfrog via forward deployment, Forge, Studio for end-to-end agent automation.
Forge used in US for post-training (e.g., financial services).
Need general models for orchestration. Enterprises have IP/physical signals (factories/tools). Open models enable deeper customization vs. closed. Run on any cloud/hardware/edge.
Work with SMEs to build bespoke apps: modify models/harness. Example: SML benefits from data-specific models.
Post-training large industry (Scale AI, Micro One). Forge: portable platform (services/training/data tools) deployed on customer infra. No data leaves; full security.
Example: ISML image scanning/defect detection. Send PhD engineers/scientists for expertise transfer with SMEs. Retrain models independently post-handover.
Exhausted open web data. Use synthetic to warm up/compress (large models teach small). Efficient start, but need human signal/feedback from experts (costly). Not enough alone.
When executives saw OpenClaw racing to the top of GitHub with massive stars, likes, and contributors, it resonated with enterprise needs. Enterprises need to automate full processes for AI gains, but OpenClaw alone isn't enough due to data problems, governance issues, lack of observability, and control. For mission-critical systems like HSBC's KYC process, deterministic gates are required that are observable and guaranteed, which OpenClaw lacks for collective productivity.
OpenClaw demonstrates autonomy for individual hackers, showing enterprises that with the right control plane, sandboxes, data sources, and access controls, agents can empower employees safely.
One user gave their agent root access to Google Docs, G Suite, Notion, Zoom, and Google Calendar. With enterprise Gmail, they summarized all conversations for their 21-person investment company, correlating Gmail and Slack. They discovered compensation discussions and a performance improvement plan (PIP), realizing the need to restrict access. Power comes from data access, but with great power comes great responsibility.
Enterprise data isn't centralized for everyone; a layer must understand data semantics to control access, e.g., preventing compensation data from flowing to engineering or HR. A context engine maps data locations with metadata restricting access. Engineering queries on comp get denied. This requires rethinking IT systems, connections, information flow, and management—agents may eliminate middle managers relaying info. Need sandboxes, RBAC (role-based access control), and restructuring customer service to reduce human info transfer.
Daniel Roberts, co-CEO and co-founder of IREN (publicly traded, started in BTC mining), from Sydney. Thesis: explosion of digital world would strain real world, so build large-scale data centers. Bitcoin mining bootstrapped cash flow; now swapping all Bitcoin for AI chips.
False dawn in 2020 with Dell MOU for AI compute, but too early—returned to Bitcoin. Real shift ~2 years ago; demand escalates monthly. Developed own data centers: find land, permits, grid connections. Flagship Texas site: 750 MW, built 4 years ago in desert—unheard of then. Traditional industry baffled; now pays dividends post-ChatGPT and OpenClaw.
Microsoft early partner; $9.7B contract late last year = 5% of capacity. Constraint is power, but IREN secured it early: 4.5 GW total (almost Bay Area's annual use). Focus: time to compute—tradespeople in West Texas trailers breaking ground, foundations, water cooling. Real-world challenges meet digital demand: supply chain stress (e.g., memory), thousands of workers in remote areas.
Like gold rush towns; hire local, support community ($1M cumulative grants: playgrounds, fire departments). Expand radius 20 miles if needed. Sites near old manufacturing with heavy electrical infrastructure; rehire, retrain locals. Workforce depleted; partnering with universities, trade colleges. Trades salaries rising: lower end ~$150K (directionally).
AI takes jobs on one side, but massive talent shortage for data centers (electrical, construction, HVAC).
100% renewable since inception: hydro in British Columbia, wind/solar in West Texas (45-50 GW available, but only 12 GW transmission to Dallas/Houston). Locate at low-cost excess renewable source, monetize into digital commodity, export at light speed. Data centers follow wind turbines/solar. Utility handles variability/duck curve via grid connection, guaranteeing 24/7 power (scarce/valuable). Backup possible to gas if needed.
Gang busters; cannot meet demand, no idle GPUs worldwide. Software improvements (50x token cost reduction per Jensen) won't slow—Jevons paradox: more compute lowers latency (e.g., ChatGPT images from minutes to 5-10s), induces more usage like added highway lanes induce traffic.
Custom silicon (Google, Amazon, Meta) emerging, tying up capacity; Nvidia has massive head start with ecosystem/standards. Safest: follow Nvidia roadmap. Desktop trend real: Dell/Nvidia powerful desktop (750GB RAM) for local open-source models like OpenClaw. Software breakthroughs distribute compute; agents, AVs, robotics compound demand.
Trump admin embraces all sources; nuclear (SMRs safer, different architecture) takes decade+ but start now: policies, capital. Tracking; inevitable. Impact: opens market, boosts US advantage. Bitcoin miners were canaries near hydro/nuclear. Unlocks distributed clean generation, accelerates Jevons flywheel.
Data center as new computer: cabling, latency, GPU hops, Infiniband/Ethernet fabric critical—every ms matters.
Elon's space data centers: hard (cost, radiation), but he's often right (maybe late).
Data egress not issue: myth of needing metro proximity busted—Texas fiber everywhere; West Texas to Dallas carrier hotel: 6ms roundtrip (immaterial).
129 job ads open; hiring aggressively.
I'm going all in.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
