CoreWeave Is a B-Tier Asset, Say VCs Lemkin and O'Driscoll

CoreWeave runs at roughly a $7B revenue rate against $10–20B of annual capex and $35–40B of debt — and venture investor Jason Lemkin calls it, along with Oracle, a B-tier asset that should trade at a discount. That was the sharpest judgment in a 59-minute 20VC conversation between Lemkin, host Harry Stebbings and Scale Venture Partners' Rory O'Driscoll, which ranged from the structure of today's venture market to Anthropic's IPO math and the inference economy.

CoreWeave fell 15–20% in the week before recording. O'Driscoll's explanation is structural, not news-driven: a neocloud that underwrites high growth with heavy debt has no way to hunker down if demand slows even slightly, so the stock swings with every data point about AI growth. Lemkin's framing is blunter — CoreWeave exists because Nvidia needs it to exist and owns no unique silicon or software, while Oracle is renting excess capacity and buying GPUs for Stargate because it cannot sell enough enterprise software.

O'Driscoll resists the tier framing but concedes the punchline: "compared to Nvidia, we're all fucking B-tier assets." His counterpoint is that second-tier names like Micron, up roughly 350%, and Nebius have delivered the returns — though he warns that putting 20% of a personal portfolio into them carries real risk.

What does this concentration mean for founders raising now?

Lemkin opened with a provocation: a company with $20M ARR, 600–700% growth, profitability and a strong team cannot get a meeting. He cited one of his own portfolio companies that raised a $3M seed and now plans never to raise again, because monthly investor updates draw crickets. Stebbings called the market's "specificity box" narrower than anything he has seen — everything outside it is "wholly unfundable."

The mechanical driver is fund size. "You got to have billion-dollar positions in everything in big boy venture today. I think billion dollar positions are nothing. There's no point in investing otherwise with large funds," Lemkin said. A $2B outcome "just isn't good enough" unless the fund owns 51%, which founders rarely accept. O'Driscoll counters with the base rate: the US tech industry produces 10 to 20 decacorns per decade, while the market is implicitly pricing 90 this year. One of those numbers is wrong, and he'll bet on 50 years of data.

The practical upshot: top-decile founders now have to meet a hundred VCs, not seven — and someone will take the bet. Seed has also inverted, with $50–100M entry prices and $300–500K allocations meaning even an 18x markup on an $18B company "doesn't move the needle."

Is Anthropic's $1.5T price defensible?

Anthropic's reported Q3 showed $16.8B in revenue, up 45% quarter-over-quarter, an 85% gross margin and adjusted profitability — yet a rumored IPO price cut from $2T to $1.5T read as a crack. O'Driscoll's bull-case arithmetic: adding about $5B per quarter takes Q4 to $20–21B and next year to $100–110B, or 12–13x forward revenue at $1.5T. "That does not feel crazy," he said.

Lemkin's worry is variance, not level. Developers aren't loyal, and he wouldn't be surprised if Anthropic's market share falls 10% by the end of 2027 while the absolute business stays excellent. His tactical advice: raise what's needed at $1.5T and move on, rather than stressing the capital base chasing $2T.

The Replit story illustrated the routing economics. Lemkin, a 14–15 hour-a-day Replit user, called the reported OpenAI drop a misread: Replit flipped its default from Anthropic to OpenAI models for equivalent quality at 30–40% lower cost, with CEO Amjad Masad confirming the price logic on stage. Router businesses like Lovable, Base44 and v0, each around $1B in revenue, face ruthlessly cost-sensitive customers.

Where's the money actually moving?

Lemkin's strongest claim: code review is the most important category in software today. AI makes generation fast and 95% correct, but getting code into production still takes an order of magnitude longer than building it. "It's built. It works. It works in dev. It looks great. How the hell do I get that thing into prod? That's where the big money is going to be made now," he said. O'Driscoll calls coding "the mother lode of LLM value" but warns with an Icarus metaphor that apps closest to the foundation models risk absorption every six months.

Modal is the market nobody sized. Its $750M Series D at a $15.75B valuation — led by Excel, per the episode — came with a run rate above $1B, up from under $100M in late 2025. O'Driscoll sizes the four or five key inference providers at $5–6B combined, versus $100–110B-plus for OpenAI and Anthropic: less than 5% of the AI economy by dollars. The risk is that neoclouds ask why they carry the capex while someone else captures developer margin.

Elsewhere, Berkshire Hathaway bought $330M of new Bending Spoons shares — Lemkin's read: "Bending Spoons is like the Berkshire version of software, right? It's like See's Candy or Coke." He also flagged the bootstrapping trap: "I almost wondered if this one and done, which I used to be a fan of, is going to be a trap in the new age. None and done is much better," citing Zapier's roughly $300M revenue and O'Driscoll's math that buying out a 30% investor at six times cash flow ties up two full years of cash.

On public comps, Lemkin questioned Anduril's $90B mark — roughly 40x revenue — against Braze acquired at $5.8B near 3x revenue and HubSpot and Monday around 2x. Harvard's endowment returned 27.1% on $68B, helped by a SpaceX stake worth about $2.2B. Stebbings expects the venture market to heat up as SpaceX distributions recycle and LPs anticipate Anthropic and OpenAI liquidity in the first half of 2027.

The two watchpoints: Anthropic's post-IPO trading, where Lemkin expects choppiness from share variance, and whether the inference layer's $5–6B-versus-$100B gap becomes the opportunity of the cycle — or a margin pool the neoclouds simply take back.