Anthropic Buys Compute From Elon & Commits $200BN to Google | Cerebras IPO | Ramp Raises at $40BN
Summary
- Anthropic’s $200 billion Google commitment and SpaceX capacity purchase turn compute scarcity into the clearest signal yet of model-market consolidation. Rory read the SpaceX deal as xAI shifting from compute buyer to seller after its data center was reportedly only about 11% utilized, potentially adding $3 billion-$5 billion annually to SpaceX’s approximately $20 billion revenue run rate. “Needs must when the devil drives.”
- Goldman’s forecast for 24x token consumption by 2030 may radically understate parallel-agent demand, but raw consumption is not the same as productive demand. Jason argued that ten concurrent agents could push the multiplier toward 250x, especially when coding agents generate ten implementations and select the best two or three. The counter-case is enterprise discipline: average engineers may not productively consume $10,000-$20,000 of tokens monthly, and quotas create “token trashers.”
- Model providers can erase prompt-sized applications, while focused enterprise workflows still have a defensible place. A $2,000-an-hour lawyer is unlikely to replace a roughly $150,000-a-year Harvey deployment with a $200-a-month Claude subscription when integrations, reliability and hallucination risk matter. Yet software without a reason to exist for agents could enter “a terminal state of decay,” with obsolescence compressing from a decade to 18 months.
- The public software tape now punishes both deceleration and an unforgiving starting valuation. Monday rose roughly 20% because it raised guidance from an already depressed valuation, while HubSpot fell about 18% after lowering guidance; Cloudflare and AppLovin remained vulnerable despite strong operating numbers because expectations were far richer. Bill Gurley’s reminder: “Price is the vector.”
- ZoomInfo is the brutal case study for how challengers can confiscate an incumbent’s growth before destroying its revenue base. Clay’s data waterfall commoditized individual providers and its agent layer captured the growth ZoomInfo needed, leaving ZoomInfo near 1% growth and guiding toward contraction. At roughly 1x revenue and 35% adjusted operating income, a take-private is conceivable—but the public-market escape route is still growth, profits and a credible future.
- Cerebras’s 20x-oversubscribed IPO is technically set up to pop, while its two-year outcome remains unknowable. The range rose from $115-$125 to $150-$160, with the $4.8 billion offering implying a $48 billion fully diluted valuation; Bill said the pricing committee was likely seeking a roughly 20% debut gain, with a retail-driven overshoot possible. The long-term wager is that faster inference earns a durable niche despite customer concentration, competition and contracts that make the future unlike the past.
- Ramp can be strategically exceptional and financially frightening at the same time. Its procurement agents and expanding software layer improve the mediocre economics of corporate cards, but roughly $1 billion of revenue against a $40 billion valuation means underwriting about 2½ years of doubling merely to approach Brex’s cited 6x multiple. Rory called that “the outer edge of terrifying.”
- The founders capable of producing venture-scale outliers are often permanently changed by the required intensity. Jason put the breakpoint around four to five years: after that, “vacation doesn’t do it anymore,” and rational people usually accept offers of $50 million, $200 million or $1 billion. Rory agreed sacrifice is real but insisted that sleep, health and perspective remain performance requirements because founders “on tilt” make worse decisions.
Deep dive
1. Anthropic is closing the gap between owning shares and merely betting on them
Harry framed Anthropic’s board-approval requirement as a material secondary-market event, citing reported marks around $200 billion-$400 billion. Rory separated legitimate primary SPVs—which invest directly and appear on the cap table—from unapproved secondary transfers of employee or investor shares.
Rory’s mechanics matter: when Anthropic refuses a transfer, a shareholder might instead contractually promise a buyer every future dollar generated by the shares. The buyer then has no cap-table position or shares, only a claim against the seller—who might be lying, lack the shares or have promised the same economics twice.
Anthropic’s documents may prohibit transferring beneficial ownership, but invalidating the arrangement at company level does not necessarily extinguish the private contract. That leaves buyers enforcing claims against sellers after an IPO, while Anthropic tries to avoid courts deciding it knowingly acquiesced to a long-running shadow market.
Jason’s pushback — worth keeping: this was a “nothingburger made up on social media,” because board consent has become standard in startup charters and Anthropic warned Memo and other investors about SPVs the prior year. His interpretation was enforcement escalation: greed persisted, so Anthropic publicly named alleged bad actors.
2. Selling compute to Anthropic recasts xAI from contender to supplier
Rory called the SpaceX agreement “needs must when the devil drives.” Elon Musk had attacked Anthropic as recently as three months earlier, but excess capacity met Dario Amodei’s shortage while Musk entered litigation with OpenAI: “the enemy of my enemy is my friend.”
The sharper inference was market consolidation. With the xAI data center reportedly around 11% utilized, Rory argued xAI/Grok was shifting from a net buyer of capex to a seller, implicitly conceding that it was not presently keeping pace with Anthropic and OpenAI as a leading-edge model contender.
Selling that capacity could contribute roughly $3 billion-$5 billion annually against SpaceX’s approximately $20 billion revenue run rate—a potential 15% lift from what recently looked like a money pit. Jason’s framing was simpler: each SpaceX business unit has a P&L, and the executive responsible for idle data centers will welcome the revenue.
The SpaceX IPO model consequently becomes unusually synthetic: Jason sketched a company moving from a roughly $15 billion run-rate business to $23 billion within two quarters as the xAI acquisition, capacity sales and then Grok are pro forma’d into the numbers. “That’s why the bankers are going to earn a couple of hundred million dollars.”
3. Google is financing a competitor that also validates its infrastructure
Anthropic’s $200 billion, five-year Google commitment embodies hyperscaler co-opetition. Jason saw a rational portfolio: Google wants Gemini to win but can keep its infrastructure productive, preserve internal competitive pressure and monetize whichever model attracts demand.
Harry estimated that Anthropic now represents about 40% of Google’s future backlog, showing how dependent hyperscaler growth has become on two private model companies. His uncertainty is important: value probably accrues to differentiated models, but the capacity to invest hundreds of billions in data centers might itself prove to be the moat.
The cited enterprise shares strengthen Google’s hedge: Jason said Gemini had moved from 27% to 40% and Claude from 21% to 48% over roughly nine months; totals exceed 100% because customers use multiple models. Google therefore participates in both fast-growing platforms, even if it would prefer Gemini itself to need all $200 billion of compute.
4. Parallel agents make 24x token growth look conservative
Jason’s immediate reaction to Goldman’s 24x token-consumption forecast by 2030 was that it sounded “way too low.” Many workflows do not need armies of agents, but those that benefit from parallelism can multiply demand before accounting for enterprise adoption outside technology.
His best example was coding: instead of building a feature once, ten agents can produce ten versions, let the model rank them and show the human the best two or three. “Why build a feature once if you can build it 10 times?” Ten agents turned his rough 24x baseline into 250x.
Harry complicated the arithmetic: raw chip performance may improve roughly 3x every 18 months, while quantization and other optimizations could produce roughly 10x more tokens per dollar every couple of years. Yet workloads progress through their own 10x steps—from chat to co-work analysis, coding and parallel agents—leaving two opposing order-of-magnitude curves.
5. Token budgets will separate 100x engineers from expensive imitation
Jason surfaced a growing CTO counterargument from teams already several releases into adoption, rather than those newly experimenting with Claude 4.7: companies may be generating far more code than they can review or ship. Running Claude Code or Codex for ten hours is not automatically enterprise-grade productivity.
Harry applied Goodhart’s law: once management measures token consumption, employees alter the variable. Token quotas can produce pointless burning to appear compliant, making consumption a worse proxy just as management needs to determine whether LLM spend should equal 2%, 5% or 10% of developer salaries.
Budget scrutiny is unavoidable. Harry noted that Anthropic was at a cited $9 billion run rate in December, yet few CIOs had budgeted for anything close to 10x that; corporate America may need to locate another $50 billion-$60 billion. Enthusiasm will therefore meet demands for measurable output.
Jason reconciled the talent debate by splitting the workforce: top engineers can become “100x engineers” and should receive every tool, while mediocre developers may consume huge token volumes for little value—the caricature being 100,000 lines of code for a blog. A survey of 30 VPs of engineering found rising spend but no convincing success heuristic.
6. Horizontal intelligence will kill features faster than it kills workflows
Harry noted Anthropic’s ten financial-agent templates and a reportedly forthcoming legal product, suggesting new domains could extend demand beyond developers. Jason would not extrapolate too quickly: Claude Design had not yet become the predicted category killer, though model features had already forced some startups into repeated reinvention.
The legal discussion explains the boundary. A lawyer billing $2,000 an hour may use Claude alongside Harvey or Legora, but is unlikely to replace a roughly $150,000 annual system with a $200 monthly subscription when briefs, DocuSign integration, review controls and hallucination risk carry career consequences.
Rory’s platform analogy favored focused vendors: Microsoft owned the operating system and Office, yet Siebel, SAP and others flourished above it; investors wrongly feared AWS Redshift would eliminate Snowflake. Claude Cowork could become an Office-like horizontal layer, while coordinated enterprise workflows still reward specialization and customization.
Rory distinguished model encroachment from agentic obsolescence. Prompt-expressible applications can vanish into Claude, while legacy marketing automation may decay because agents do not need HubSpot, Marketo or Salesforce templates. Lovable and Replit cannot lag the underlying models by a week; software that once aged over ten years might now age in 18 months.
7. Guidance and starting price explain the software tape’s apparent chaos
Monday and HubSpot were the clean comparison: both are decelerating and early in their agentic transitions, yet Monday rose about 20% after genuinely raising next-quarter guidance while HubSpot fell roughly 18% after lowering it. Raising guidance at least tells investors, “we’re not going to zero.”
Cloudflare delivered what Rory described as a roughly mid-30s quarter and cut about 20% of its workforce, but the stock declined as investors wondered why a good company needed such a reset. AppLovin’s roughly $7 billion run rate did not protect it either; Bill.com, meanwhile, bounced after layoffs and a buyback.
Rory separated operational direction from valuation. An investor might accept a messy transition at 4x revenue but reject it at 15x; Cloudflare and AppLovin entered the paradigm shift richly priced, whereas Monday sat below 2x revenue with substantial cash and at one point approached 1.5x cash.
Jason’s broader call was that the “SaaS apocalypse” may be past, but fear of zero terminal value remains. When surrounding budgets accelerate while a vendor decelerates, those lines do not create a forgiving setup: “If you’re not accelerating, you’re going to be destroyed.”
8. Clay did not destroy ZoomInfo; it stole the growth ZoomInfo required
ZoomInfo was growing around 1% and guiding toward negative growth after winning the pre-AI sales-data market. Jason estimated Clay at roughly $200 million-$300 million against ZoomInfo’s billion-plus scale, yet concluded that Clay and peers had taken “all of ZoomInfo’s growth away from it.”
Bill’s mechanism was more specific than “AI disruption”: Clay’s waterfall lets RevOps compare five or six data providers, turning any single dataset into a commodity. Clay then layered agents and a stronger AI narrative onto that pre-LLM wedge, pulling value away from the underlying providers.
At roughly 1x revenue, 35% adjusted operating income and approximately neutral customer growth, ZoomInfo resembles a classic take-private—but Bill warned that “PE buys it” is often lazy analysis. The alternative rerating formula is explicit: 30% growth, profits and a future-facing story might earn 5x-6x, never the former 20x.
9. Cerebras’s IPO pop is much easier to forecast than its business
Demand was reportedly 20x the offering, prompting Cerebras to lift its range from $115-$125 to $150-$160. The $4.8 billion raise would value the company at approximately $48 billion fully diluted; Bill said bankers would not make that change without high confidence in the book.
The desired outcome is a roughly 20% first-day pop, though retail enthusiasm could create a Figma-style overshoot. Bill cautioned against treating such volatility as fundamental performance: Figma priced near $35, briefly traded around $100 and later fell significantly below its offer price.
Jason saw the right timing, partners and apparent backlog, including OpenAI and Amazon, but stressed that production history remains short and customers will hedge across Cerebras, NVIDIA, Groq and other solutions. Historical revenue was concentrated among a few UAE customers, while the forward story depends on contracts unlike that past.
Cerebras sells speed: “How much would you have to be paid to have a slower internet?” Eric cited his company Tabula as a concrete use case requiring real-time responsiveness. Against NVIDIA’s cited $5.5 trillion value, $48 billion offers an understandable at-bat—but Jason might still “take my profits.”
10. Cerebras is also a rare specimen of actual venture creation
Harry initially cited 20% ownership, but Rory corrected him from the S-1: Foundation, Benchmark and Eclipse each held roughly 8%-9%. Maintaining that stake across eight or nine years in a capital-intensive semiconductor company was itself an exceptional outcome.
Foundation earned Harry’s “jealous” admiration, which Jason recast as impressed, by helping incubate the company around 2016, before the category was obvious. Cerebras survived being early in 2021-22, found business in the UAE, withdrew an earlier IPO attempt, then secured OpenAI and Amazon commitments before returning to market from strength.
Harry called that “real venture capital,” and Jason agreed: finding Andrew Feldman, developing the relationship, seeding the company and enduring the hard years—not using a large firm’s brand to muscle into a later round. Feldman’s persistence was equally decisive; merely very good founders would have quit.
11. Ramp’s product expansion is compelling; its multiple assumes perfection
Rory framed corporate cards as broad but economically limited: interchange revenue requires customer rebates, leaving only acceptable contribution margins. Ramp improves the business by adding ACH, software and procurement agents that can analyze spending, contact suppliers and negotiate pricing.
Jason’s buying criterion became deliberately concrete: after rebuilding his finance stack, he would choose between Brex and Ramp partly on which offered the best agents. Procurement’s “moronic back and forth,” fake pricing and ritual 10% concessions are precisely the workflow he wants agents to remove.
Price remains separate from strategy. Ramp was cited around $1 billion in revenue and raised at $40 billion, versus a roughly 6x Brex comparison at lower growth. Even doubling from 1 to 2 to 4 to 8 requires about 2½ years before Ramp reaches that multiple—“the outer edge of terrifying.”
Lime supplied the more conventional recovery story: after surviving an exceptionally hard operating model, it is preparing for an IPO and, according to Harry, dominates large parts of London. The panel’s reaction was less valuation analysis than recognition: “Oh my god, they’re alive.”
12. Memory stocks require a cycle forecast, not an aversion to fivefold gains
Asked about Micron and SK Hynix after roughly 5x appreciation, Rory rejected “it already went up” as sufficient analysis because earnings could still make the stocks look cheap. The real variables are the remaining duration of AI capex demand and how quickly manufacturers add DRAM fabrication capacity.
The recurring danger is synchronized reversal: demand slows just as new fabs arrive, crushing pricing. Rory pointed to the post-COVID laptop hangover and the sector’s 2022-23 weakness; when producers earn 50%-60% net margins, “the temptation to build a fab just becomes huge.”
His 2028 image was illustrative, not a precise forecast: Samsung and SK Hynix could be digging holes just as Anthropic begins cutting orders. Rory had not committed his own money, underscoring that a coherent framework is not yet a developed position. “Ain’t capitalism great?”
13. Extreme outcomes require intensity, but tilt still destroys value
Rory agreed that meaningful success entails sacrifice—time, alternative lives and sometimes relationships—but rejected mental deterioration as a badge of honor. When stress eliminates perspective, decision quality falls; sleep, health and coping mechanisms are part of sustaining intensity, not concessions to it.
Jason contrasted two startups. After selling the first for $50 million in 12½ months—following pulled financing, personal payroll funding and a full-recourse loan against his house—he felt normal within a month. Five years building the second through repeated near-bankruptcies and the financial crisis “permanently rewired” him.
His breakpoint was roughly four to five years: vacations, runs, watches or yachts no longer restore the old self. Daniel Dines’s line carried the emotional cost—“when the lights go out at the end of the day, it’s very lonely in my head”—because the defining burden is intensity, not merely hours.
Jason therefore tells founders to accept $50 million, $200 million or $1 billion offers when proceeds exceed roughly 3x capital raised—unless they instinctively reject the exit. Normal people sell; outlier builders self-select by continuing. Jason’s final formulation was that the claim is both “real and rage bait” because its truth annoys people, and Rory agreed.