Fireworks

Instinct Worth 30BN? | Microsoft Cuts 27,000 Jobs While Meta Taps Equity Markets for $95B CapEx

Oct 29, 2026 · 1:00:17

Harry Stebbings, Jason Lemkin and Rory O'Driscoll argue Anthropic's S-1 shows a company investors are still underestimating, with $16.3 billion first-half revenue and $110 billion of cash, though 80% of revenue comes from models under six months old. They break down Microsoft's buyout offers to 27% of its 223,000 employees, warning 2027 CapEx depreciation will force more AI-driven layoffs, and Meta raising $95 billion of equity for its spending spree. Jason argues Chinese open-weight models are finished in the US after Alibaba's Quen was caught uploading data to servers in China and Singapore. They debate Instinct at $30 billion as the 'super app of the West', whether OpenRouter's boom is a moment in time, and Jason's pitch to put 20% of the fund into OpenAI at $1.25 trillion.

Inference Will Be $10TRN Market: What No One Sees | Software Investors are F***ed | Naveen Rao

Oct 24, 2026 · 1:13:26

Naveen Rao, founder and CEO of Unconventional AI, argues that energy will be AI's biggest bottleneck and that the inference market will reach $10 trillion within 10 years. He explains that physical efficiency gains have stalled, so value comes from co-designing model, architecture, and hardware—why chips will not commoditize and Nvidia keeps the margin. He tells Harry Stebbings that pure software investors are 'kind of fucked' as SaaS fragments, with AI-plus-human outcome delivery replacing it, and that the bottleneck has shifted from creation to verification. Rao predicts robotics will be 100x bigger than agentic coding, open models will carry 80% of token traffic, and traditional VC metrics like market validation no longer fit deep tech.

Who Wins in the Next Decade of VC | Margin and Price Matter Less in a World of AI | Sameer Dholakia

Oct 19, 2026 · 1:19:51

20VC's Harry Stebbings hosts Bessemer partner Sameer Dholakia, who argues AI is a bigger supercycle than cloud and the internet, with intelligence a bigger market than oil. He recounts Bessemer's Anthropic Series D at $15B pre-money—the firm's lowest-voted deal ever—plus bull cases for Legora at $1.8B and Fireworks at $250B, holding that open and closed models both win. He contends capital is a moat, T2D3 growth is no longer enough, and defends 10-15x NTM entry multiples. He explains Bessemer's new $5.75B funds ($1.75B seed, $4B growth), why 5x returns remain possible, and why venture is splitting into mega-platforms or specialists. He predicts companies will stay private longer with an IPO unlock next year and in 2028, and credits SendGrid's 'teammates beat colleagues' culture.

Horizontal vs Vertical Agents: Who Wins? | Model Routing: Commodity or Increasing in Value?

Oct 17, 2026 · 1:03:06

Harry Stebbings hosts Harrison Chase, CEO of LangChain, who argues agents are hard to build because models are non-deterministic, and value accrues to whoever controls context, harness and evals. Chase explains why model routing belongs in the harness for vertical agents but in gateways for horizontal tasks, why agent memory is unsolved after LangChain killed its memory product, and why vertical agents should start at the harness before post-training. He says inference and coding agent demand is far bigger than expected, that LangChain spends ~30% of engineering salary on coding tools — a figure he ties to Anthropic's valuation — and that open models mostly serve simple tasks, not core loops. He closes with his bull case for a $9BN Jev round and how Benchmark and Sequoia help founders.

Crusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong

Oct 3, 2026 · 1:05:02

Crusoe CEO Chase Lochmiller tells Harry Stebbings that data centers lower energy prices, use almost no water, and that most moats in AI are an illusion. He explains Crusoe's pivot from Bitcoin mining to AI infrastructure, accelerated by ChatGPT's November 30, 2022 launch, and how Crusoe built Abilene's first 200-megawatt buildings in one year when 34 competitors bid two and a half years, vertically integrating electrical manufacturing to cut a 100-week power distribution lead time to 28 weeks. He names energy and skilled labor as the true bottlenecks, concedes 50% of planned data centers may never be built, and frames Crusoe's three products—data centers, GPUs, and tokens—through an 'AI super major' analogy to Exxon's vertically integrated hedge. On GPU depreciation, he argues six-year cycles underestimate H100s, whose rates today exceed launch pricing, defends take-or-pay rental contracts, and predicts closed models capture more dollars while open models generate more tokens.

Inside Sequoia's Investment Committee | How the SpaceX & Citadel Deals Went Down | Julien Bek

Aug 24, 2026 · 1:28:28

Julien Bek, a Partner at Sequoia Capital, says Sequoia wins by hunting with conviction, not waiting for inbound calls, and shares founder-reading lessons from Don Valentine, Doug Leone, Pat Grady, Alfred Lin, and Shaun Maguire. He highlights Leone's 'worst reference' question, Lin's warning not to mistake an outlier operator for an outlier founder, and Maguire's ELO method for calibrating references. Bek also explains how asking 'why' five times exposed a fraudulent founder, how his read changes by country (French/German references get 1-2 extra points; Americans get subtracted), and why he missed Anton Osaka of Lovable. On AI, he argues agents become the new customer, AEO parallels SEO, and the next trillion-dollar company will be software masquerading as services. He tells how his mother invested in Revolut at a ~$200M valuation, now worth over $100B.

The AI Bubble WILL Burst | Should we be fearful of Chinese Open-Source | Jerry Murdock

Aug 22, 2026 · 1:10:09

Jerry Murdock, co-founder of Insight Partners, warns the AI bubble may burst by March if the Iran war continues, citing credit market complacency and Japan’s $1 trillion Treasuries. He predicts at least half of Neo Clouds disappear within 36 months, favors Fireworks over Base10, and argues open-source models plus ASIC chips will drive customization while frontier labs retain complex tasks. On security, he calls sandboxes the most underrated layer, naming Docker and E2B as leaders, and expects OpenRouter’s 5% markup to be disrupted by inference exchanges. He adds private equity is highly at risk to any financial dislocation, and in quick-fire bets Nvidia exceeds $10 trillion in five years while picking Meta as the Mag 7 short.