Anthropic Raises $30BN at $900BN Price | SpaceX Files S1: How Does it Trade | Cerebras Smashes Day 1
Summary
Anthropic’s proposed $30 billion raise above $900 billion may still be “the best value in the venture universe” if ARR remains the right yardstick. At roughly 18x June revenue, with 10x growth and little remaining IPO or existential risk, it compares favorably with early-stage rounds at 20–50x ARR. Selling about 3% also buys another year in a “big-ass balance-sheet war” where high-end compute commitments run into tens of billions.
Salesforce’s $300 million annual Anthropic bill is simultaneously enormous and ordinary. Spread across 20,000 developers, it is about $15,000 per engineer, $1,200 monthly and 4% of engineering spend—almost exactly the panel’s survey average. Yet projected trillion-dollar token markets require roughly 20% of engineering payroll and 5–7% of knowledge-worker wages, making Salesforce perhaps “only a quarter of the way” there—or exposing an overestimated TAM.
Public SaaS has time to recover, but the 2021 valuation regime is permanently gone. Datadog crossed $4 billion alongside an all-time-high ARR, a first $1 billion revenue quarter and 32% growth; Figma’s NDR reached 139% as growth approached 50%; Atlassian and Twilio also reaccelerated. The panel’s memorable reset: “They’ll never get back to ’21 prices, and I’ll never be 21 again”—great execution can restore normal multiples, not the old speculative veneer.
Figma can monetize the software-building boom, while Wix shows what terminal expectations look like. Figma missed at least a $500 million opportunity by failing to turn approved designs directly into production software, but its new agentic design tools could extract substantially more from its base. Wix, down 45% since its repurchase and near 1x revenue despite Base44 reaching $150 million ARR, is being squeezed by both vibe-coded websites and Shopify; “can’t get worse” still does not mean it gets much better.
The compute trade remains intact because scarcity is outrunning every visible warning about eventual oversupply. Nebius grew 684%, Cerebras priced at $185 and popped 68% to break above $300, and the panel sees “no signs that there’s a short-term crash coming.” Permitting delays may paradoxically prevent ruinous overbuilding, but the entire chain ultimately depends on corporate software converting infrastructure into enough paid token consumption.
SpaceX’s $1.75 trillion, $75 billion IPO could become the ultimate test of retail-driven price discovery. Its S-1 may mostly describe the old SpaceX and Starlink, while barely reflecting xAI and excluding the signed Anthropic and pending Cursor deals; the missing story will therefore be sold through the roadshow. Jason forecast a run toward $5 trillion, Harry a move to $3 trillion, while Rory bet it would finish its first month below $3 trillion: “This is a meme and this is a casino.”
AI’s capital boom is colliding with a political backlash its own leaders helped create. Eric Schmidt was booed, while Meta, Cisco, LinkedIn, Intel and Standard Chartered announced thousands of cuts; Jason fears the laid-off will carry a “double scarlet letter” because nobody will rehire them. Harry argued that scientists who are brilliant at AI but “utter morons” at politics will be eaten alive unless the industry addresses job loss rather than acting surprised by the anger; Rory separately said public sentiment toward AI has sharply worsened.
Deep dive
1. Anthropic can sell 3% because compute, not valuation, is the binding constraint
The opening news paired Andrej Karpathy’s move to Anthropic with talks to raise $30 billion above $900 billion, nearly tripling February’s $380 billion price. The named investors were Greenoaks, Sequoia, Altimeter and Dragoneer.
Rory’s framing: since Anthropic’s roughly $150 billion round, investors have effectively been buying a post-IPO-quality asset without meaningful IPO or disappearance risk. Against private companies with $10 million ARR priced at 20–50x, Anthropic at roughly 18x June revenue and far higher growth has repeatedly been “the best trade out there.”
That conclusion remains conditional: if ARR is still the correct proxy despite compute costs and lower gross margins, it is “the best value in the venture universe”; if everything must ultimately reduce to discounted cash flow, the comparison is less clean. The panel noted that growth investors still largely price ARR without the margin discounts common before 2021–22.
Anthropic’s answer to why it accepts an apparently investor-friendly price is dilution math: $30 billion over $900 billion is about 3% surrendered to derisk another year. Jason said Anthropic was committing roughly 6.5 gigawatts that year, with a rough $40–50 billion total cost per high-end compute gigawatt. This is a “big-ass balance-sheet war,” even when hyperscalers fund much of the build.
2. Anthropic and OpenAI are running opposite financing playbooks
Jason contrasted Dario’s low-drama approach with Sam Altman’s valuation maximization. Anthropic appears willing to accept a “70% deal” completed within 72 hours; OpenAI pushes until demand exceeds supply by exactly one dollar, consistent with its need for effectively infinite capital.
The last Anthropic round was described as simple: commit by email, then wire $30 billion in cash. The contemporaneous OpenAI financing was portrayed as a $110 billion structure with only $20 billion clearing immediately, additional Amazon capital contingent on an IPO or AGI, and SoftBank funding dependent partly on borrowing.
Anthropic may not need another private round if its stated November IPO timing holds, but the panel rejected the idea that continued fundraising is unhealthy. During hypergrowth, fresh equity funds both capex and expansion; “when you stop having to raise, that’s a disaster” may mean the growth story and capital need have ended.
3. Salesforce’s $300 million token bill is the market’s most important datapoint
Marc Benioff said Salesforce spent $300 million on Anthropic tokens, almost entirely for coding. Across 20,000 developers within 83,000 employees, that is roughly $15,000 per developer annually, or $1,200 monthly—about 4% of Salesforce’s stated $5.8 billion engineering spend and modest beside a roughly $500,000 fully burdened developer cost.
A survey across about 40 portfolio and external companies found average monthly AI spend of $1,200–$1,300 per developer, with a lower median and a token-maxing tail. Salesforce is therefore “in the strike zone of normal”; the headline is huge only because the company employs so many engineers.
Yet this is probably Salesforce’s largest external vendor line item apart from something like rent, created from virtually nothing in two years. That makes $300 million both “eh” relative to payroll and an explanation for Anthropic’s extraordinary rise: the same product can be table stakes for the buyer and transformational revenue for the seller.
4. A trillion-dollar token market requires tokens to replace wages, not software licenses
The panel’s rough four-year case for $1 trillion of combined token revenue implies capturing around 5–7% of every knowledge-worker salary and 20% of engineering wages. At 1% of R&D spend, tokens disappear into noise; at 5%, “that’s a layoff”; at 20%, they consume one-fifth of engineering payroll.
Worldwide software revenue was estimated at $1.2 trillion, with R&D near 20%, or $240 billion. Even taking 20% of that pool produces only about $50 billion, so OpenAI and Anthropic cannot justify trillion-dollar projections merely by cannibalizing conventional software budgets—they must eat a substantial share of broader operating expense.
The fork is stark: Salesforce remains at $300 million, implying the token TAM was overstated and valuations correct sharply, or it quadruples usage and Benioff eventually announces a billion-dollar bill. Because Salesforce is probably ahead of most enterprises, today’s evidence does not yet resolve which outcome wins.
Jason supplied the efficiency bear case: his small SaaStr operation runs 21 agents, three autonomous, for about $2,000 monthly in direct AI costs. Models and users may improve fast enough that universal adoption requires “a half or a third or a quarter as many tokens as we think.”
5. Klaviyo shows that agentic work can be mandatory without being expensive
Klaviyo co-founder Andrew Bialecki reportedly requires every employee near product—including product and design staff—to commit code and use AI or agents. The company built a custom harness to manage model behavior, making agentic work an operating requirement rather than an optional productivity tool.
Backstage, Bialecki was asked to estimate the token cost of running two autonomous AI executives, one for marketing and one for customer success. He guessed roughly $250; the actual direct agent bill was $257. His lesson was that even a company operating agentically across the organization need not fear runaway token consumption.
Jason became more constructive on Klaviyo’s leadership but retained the external-product objection: internal AI excellence does not automatically produce a competitive customer-facing agent. Public software leaders have had about 18 months, including since Claude 4, to respond while small startups out-hustled incumbents; 2027 improves only if leaders now ship the best agents in their categories.
6. SaaS can reaccelerate, but it will be valued as an adult industry
Datadog’s first $1 billion revenue quarter grew 32% and coincided with an all-time-high ARR and a crossing of $4 billion; Figma accelerated for a second consecutive quarter with 139% NDR; Atlassian moved above 30%, aided by Rovo; and Twilio returned to roughly 20% growth. These are operating recoveries, not merely multiple expansion.
Jason’s revised view is that iconic, founder-led companies may have “a little more time than we thought” because most buyers do not live at San Francisco’s technological frontier. Monday beat and bounced, while HubSpot warned that Q2 would be harder and was punished: public markets still demand, “Show me the growth.”
Rory rejected comparisons with 2021 peaks: “They’ll never get back to ’21 prices again, and I’ll never be 21 again.” Every cycle grants one sector a youthful period when investors value possibilities; after that veneer disappears, revenue, growth and cash flow govern permanently.
The resulting range is narrower: an exceptional Datadog might command 17–18x sales, a good Figma perhaps 6–10x, and weak software around 3x. SaaS can remain an excellent business, but AI companies are one or even two orders of magnitude larger and earlier, so attention will not rotate fully back.
7. Figma missed the code-generation race but still owns valuable workflow
Jason’s self-correction was precise: he remains convinced Figma Make was the worst vibe-coding product he used and that management left at least $500 million behind by failing to build a Replit-level competitor. What “Limited Lumpkin” missed was the obvious offset—AI is creating a software explosion, and Figma sells picks and shovels to software builders.
Figma can now let an agent inspect a design and update its workflow or user journey, bringing vibe-style iteration inside the core product. At Figma’s scale that is nontrivial, and Jason thought these capabilities could plausibly extract another 50% or more revenue from the installed base.
Lovable threatens the upstream workflow by letting users describe and operate a working prototype instead of drafting a static design, though Jason judged its professional design capabilities too limited to replace Figma today. The danger is less current substitution than customers learning to route around Figma entirely.
The missed insertion point is painfully visible: both Replit and Lovable invite users to upload Figma files because their ideal customer wants to convert approved designs into production. Figma should have owned the button that says, in effect, “push into full production prototype and it just works.”
8. Wix is approaching terminal value after spending its optionality
Wix was down 45% since its repurchase, at roughly a $2.2 billion market value and later described as trading near 1x revenue, even as Base44 reached $150 million ARR. Public markets are treating the pre-AI business as terminal and doubting that Base44 can grow fast enough to offset it.
Jason saw two attacks: anyone not deeply tech-phobic can now vibe-code a better bespoke website in minutes, while Shopify destroyed Wix, Squarespace, WooCommerce and BigCommerce as credible commerce alternatives. Merchant services and e-commerce had supplied Wix’s growth, making Shopify’s success especially damaging.
Rory kept the counter-case alive: if Wix converts its large low-end SMB base to Base44 while the acquired product compounds from $100–$150 million, aggregate growth could eventually turn. At 1x revenue, a high-margin, relatively sticky product is nearing terminal value unless churn is truly catastrophic—but “can’t get worse” is not the same as “will get a lot better.”
The buyback sacrificed option value. When technology is changing this quickly, another billion dollars could fund a decisive acquisition; instead Wix optimized the near-term share price and bought before a further 45% fall. Buybacks may placate activists, but those activists will not remember requesting one when the stock subsequently collapses.
9. One founder-led incumbent may still turn an installed base into distribution
Jason predicted that one “downbeat” software company will become “upbeat”: a founder will lock the best 50 people in a room, build a superior prosumer AI product, and sell it aggressively into hundreds of thousands of existing accounts. The startups move faster, but they lack that distribution.
HubSpot was the clean example. With roughly 300,000 customers, a Breeze AI SDR as good as the startups’ products might sell into 150,000 accounts. Canva’s 2.0 effort reflects the same possibility, though the panel could not identify in advance which incumbent will execute.
The ceiling remains lower than the old dream. Rory could not name a strategy that reliably returns a mature company to 30% growth and 5x revenue; many are “managing decline.” Installed bases can create valuable normal companies, but only founders willing to concentrate talent and cannibalize the old workflow have credible upside.
10. Compute scarcity is saving infrastructure vendors from commodity economics
Nebius grew 684% and accelerated again. The call is binary: if compute remains scarce, Nebius and CoreWeave remain great businesses; if capacity becomes plentiful, they become commodity suppliers, and the most overleveraged operators go bust.
The paradox attributed to Gavin Baker was that slow permitting may “save us all from ourselves.” If industry builds $1 trillion of data centers just as projected token revenue falls to $500 billion, economics implode; if bureaucracy permits only half the planned capacity, compute stays scarce and existing owners retain pricing power.
Hyperscalers and model companies are spending roughly $750 billion–$1 trillion annually, with perhaps 50% flowing to Nvidia, 10% to power and 10% to networking. That tide has reignited memory, semiconductors and Cisco; Jason’s road-trip observation was that every technology category except traditional software is “on fire.”
Rory’s unresolved seed decision captured the duration risk: he passed on an excellent compute team because the trade was already three years into its capex boom and would require $500 million–$1 billion, yet the founder’s response “won the argument.” Public investors can trade scarcity; venture investors must underwrite capital markets years ahead.
11. Cerebras validates exceptional AI IPOs, not the entire backlog
Cerebras moved from an early $110–$120 indication through $150 to a $185 IPO price, the maximum available without refiling, then jumped 68% and broke above $300. It was described as the biggest US technology IPO since Snowflake.
Rory called it an “N of one”: a differentiated semiconductor and inference product arrived exactly when those categories exploded, with OpenAI as the marquee customer. Jason cited a $24 billion backlog in his more optimistic framing. Two years earlier the company could not complete an IPO; timing, positioning and risk appetite changed everything.
The read-through is positive for companies at Cerebras’s level or above—especially SpaceX—not proof that sub-Figma issuers can list successfully. Rory would not buy Cerebras blindly at $300: from the first-day closing price, IPO base-rate returns are negative across one, six, 12 and 24 months, even when the underlying company may endure.
12. SpaceX’s S-1 may describe a company that no longer exists
SpaceX set June 12 for a proposed $75 billion raise at a $1.75 trillion valuation, which would make it the largest IPO ever. The panel expected extraordinary demand but noted that pricing begins near 100x revenue.
Jason expected the filing to show the December-era SpaceX and Starlink: leaked figures of roughly $15–$18 billion revenue, 20–30% growth and positive EBITDA, with capex still needing scrutiny. That is a bounded, understandable company—and increasingly unlike what investors will actually buy.
The February xAI combination contributes perhaps half a quarter of minimal revenue and enormous burn; the signed Anthropic transaction will not yet appear in the financials, and the pending Cursor acquisition is not closed. Bankers must narrate the new entity and these acquisitions through the roadshow.
The bull setup is therefore mood-dependent: “the most exciting company on the planet” is arriving when markets want excitement. If investors suddenly demand cash flow, the story changes; Harry characterized it as Starlink’s growth engine with a CoreWeave-like compute business attached, not an Anthropic-like model company.
13. Retail can move SpaceX violently without making it another GameStop
Roughly 30% of the offering was expected to go to retail, and Harry said he planned to put $2,000 into it from his phone. Jason argued that rockets should excite Robinhood and GameStop traders, but a $75 billion float is not thin, and a $1.75 trillion starting value constrains the arithmetic.
Institutions create the opposing force: if an investor buys expecting 40% over 12 months and receives that return on day one, selling is rational. Rory’s hypothetical was a $10 billion allocation returning to market immediately after hitting its internal target, turning the pop into additional supply.
The bets exposed three time horizons. Jason forecast $5 trillion; Harry expected a move to $3 trillion; Rory said he was comfortable betting that it would be below $3 trillion at the end of the first month, while separately allowing that it could trade up threefold during 2026. Facebook’s weak 2012 debut remained the warning that even a generation-defining company can be overpriced.
14. OpenAI’s YC token offer is both financing and a capacity signal
Sam Altman offered every startup in the current YC batch $2 million of OpenAI tokens in exchange for equity, evoking DST’s early batch-wide investments. Rory read it as a smart attempt to recover developer “hearts and minds” after Anthropic stole several marches.
The grant could anchor valuations near the $100 million OpenAI price, shrink cash rounds and cut conventional VC ownership from today’s 5–6% toward 2–3%. A founder can reasonably ask, “I got $2 million at 100—why would I take another four at 50?”
Jason’s sharper insight was that “tokens are marketing”: startups can fund generous usage and customer acquisition rather than merely engineering. If future grants rise toward $10 million, founders can focus on shipping “the best 5.7 Codex product” without first-year token anxiety.
Capacity determines the economics. At 150 startups, $2 million each is $300 million per batch, or about $1.2 billion across four annual batches; at 18x, constrained tokens carry a $20–$30 billion valuation opportunity cost. Jason’s conclusion was that OpenAI has surplus tokens and Anthropic does not; Harry agreed, while noting that it was a bet rather than an established fact.
15. AI’s legal victories will not protect it from the political backlash
Harry said the panel’s prior call on Elon Musk’s OpenAI suit held: it was dismissed on a technicality. Rory explained the relevant issue as the statute of limitations: because Musk discussed for-profit conversion around 2016–18, the claim that he discovered the alleged fraud only in 2023–24 was implausible. Harry predicted that an appeal would go nowhere.
Jason nevertheless rejected the clean “Sam took no consideration” narrative, pointing to OpenAI’s venture fund and Altman’s alleged carry, and speculating that this complexity may relate to his firing. Rory’s rebuttal was that any indirect economics were tiny beside the equity Altman could openly have requested, but complex good intentions now give Musk endless investigative leverage.
Public sentiment is deteriorating faster than the industry admits. Rory recalled that students once applauded ChatGPT because “we’ve all cheated for the last year”; three years later Eric Schmidt was booed. After years of warnings about extinction, unemployment and rising electricity costs, leaders should not be shocked that people dislike them.
Meta’s 8,000 cuts, Cisco’s 4,000, LinkedIn’s 875, Intel’s 16,000 and Standard Chartered’s 7,800 “job reductions in favor of the machines” make the politics concrete. Jason expects a “double scarlet letter” and ultimately thousands of compensatory hires per tech leader; Rory’s hedge is crucial: if AI replaces only 5%, not 20–50%, companies may discover they cut too deeply and rehire anyway.