Anthropic's Raise & What It Means for Potential IPO? Mag7: Google & Amazon Up, Meta & Microsoft Down
Summary
- Alphabet was the clean winner in “the most aggressive quarter in American capitalism,” but the larger call was that the top of the distribution is pulling away. The Mag 7 cohort produced roughly $540 billion of combined revenue while the panel put AI CapEx at $700 billion; Alphabet’s cloud backlog nearly doubled to $462 billion, search remained intact, and its cloud business accelerated. This is “leaning in like you’ve never seen leaning in before.”
- The hyperscalers’ AI growth is real, yet much of it still comes from supplying and distributing the products of two private model companies. Google, Amazon, and Microsoft sell compute to the labs and then resell their tokens, while the labs retain the core IP; the open question is “when the music stops, who has a chair with a trillion dollars on it?” Gemini token production rose from 10 billion per minute in Q4 to 16 billion in Q1, versus roughly 10x growth at Anthropic over the same period.
- Microsoft has made AI both its only meaningful growth engine and the central risk to its valuation. Excluding Copilot and Azure’s AI contribution—with the panel noting that favorable revenue gets reallocated—the rest of Microsoft was described as flat to slightly down; AI ARR was put at $37 billion against prospective CapEx of $190 billion. Jason believes management has modeled the exposure “to the last decimal point,” but Harry’s warning survives: everyone at the top of a CapEx cycle thinks the investment is intelligent.
- Meta’s earnings beat could not compensate for an AI plan whose return still cannot be modeled, while Amazon offers clearer application-layer leverage. Meta delivered $56 billion of revenue and $10.44 of EPS versus $6.67, then raised CapEx guidance to $125-$145 billion without a clear attributable revenue stream. Amazon’s $181 billion of revenue, $37 billion at AWS, Anthropic alignment, and fastest AWS growth in 15 quarters make it the panel’s weakly held buy; Microsoft was the sell.
- Palantir is monetizing the board-level need to “do something in AI” at a scale point solutions cannot reach. RPO rose 134% to $4.45 billion and its Rule of 40 score hit 145%, while its government heritage lets it credibly sell $10 million-to-$100 million enterprise transformations rather than $200,000 pilots. At a $349 billion market cap it is “priced to more than perfection,” although two years of doubling would make today’s valuation look much less extreme.
- The SaaS apocalypse did not end, but Atlassian and Twilio showed two credible routes out of it. Atlassian rose 29% after successfully monetizing Rovo into its installed base, while Twilio gained 20% as AI builders generated new API demand and potentially drove customer growth near 40%; Five9’s 23% rally on 9% growth did not impress Jason. The durable winners must demonstrate relevance, controlled dilution and retention—and ideally attract new customers, not merely upsell old ones.
- Anthropic’s opportunity is enormous, but its stated revenue pace is difficult to reconcile with how little many productive agents currently spend on tokens. Jason’s two autonomous marketing and customer-success agents cost only $254 for a month, with the marketing agent consuming $94.27, while Jason’s framework needs mature engineering organizations to spend perhaps 20%-30% of salary dollars on tokens rather than 5%. Even so, a proposed $50 billion raise at roughly $900 billion buys critical optionality: “There is no such thing as too much cash on your balance sheet.”
- AI is beginning to reward hands-on executives and eliminate managerial layers that cannot ship. Coinbase’s roughly 14%-15% reduction was framed as “build or go”: managers must also produce, CMOs should be able to launch campaigns through agents, and a rare AI-native SDR may be worth $250,000 while the old $60,000 research role disappears. Rory resisted Jason’s absolutism but accepted the mechanism—if hands-on leaders outperform, “economics is Darwinian,” and today’s minority becomes the operating norm.
Deep dive
1. AI turned the Mag 7 quarter into a capital-allocation arms race
Harry framed the week as the earnings “Super Bowl”: roughly $540 billion in combined revenue alongside $700 billion of AI CapEx. These were not defensive incumbents protecting mature franchises; the largest companies on earth were simultaneously accelerating and investing ahead of demand.
Borrowing Evan Armstrong’s phrase, Harry called it “the most aggressive quarter in American capitalism.” The largest five or six companies were growing near 20%, some businesses were advancing 30%-40%, and CapEx was rising 50%-60% until it consumed much of free cash flow.
The structural message was sobering as well as bullish: “This is the top of the distribution pulling away.” Unlike a conventional disruption cycle where newcomers attack slow incumbents, five of the seven largest public companies—and six including Nvidia—are refusing to concede the next platform.
2. Alphabet won the quarter while Gemini lost ground to the private labs
Alphabet’s cloud backlog nearly doubled to $462 billion, approximately its entire stated 2025 revenue, while cloud growth accelerated at enormous scale. Jason’s reaction was simple: “It once again makes you wonder why you invest in anything else.”
The feared search collapse has not materialized economically or in advertising. Jason’s own SaaStr search traffic was up 60% to an all-time high, and Google’s dilemma is no longer choosing between its cash cow and the future; it is deciding how to allocate scarce TPUs among itself, customers, partners, and platforms such as Replit.
Harry’s less comfortable interpretation: hyperscalers are growing by selling compute to private LLM companies and then buying their tokens for resale through existing distribution. “The five largest market cap companies on the planet are effectively working for these two privately held companies,” while those companies own the core IP.
Gemini production increased from 10 billion tokens per minute in Q4 to 16 billion in Q1. David later contrasted that with roughly 10x growth at Anthropic. Benchmarks may suggest Gemini, Grok, and open models are close; actual developer choices in coding continue concentrating on OpenAI and Anthropic because of either the models or their harnesses.
3. Microsoft made AI both its growth engine and its valuation risk
Harry’s load-bearing statistic was that, after removing Copilot and AI-driven Azure growth—with a caveat about favorable internal reallocations—Microsoft’s remaining business was flat to slightly down. “Without the AI initiative, Microsoft the corporation is flat revenue.”
The panel placed Microsoft’s AI ARR at $37 billion and prospective CapEx at $190 billion: an aggressive investment far ahead of present revenue. If it proves excessive, Microsoft can throttle spending, live off its existing franchises, and endure a digestion period; it is not carrying the debt-funded existential risk of a weaker company.
Jason trusts the sophistication rather than the infallibility of the bet. Management has run sensitivity analyses, understands what risk can be shifted to CoreWeave or NVIDIA, and can scale back—but Harry’s pushback was that every company at the top of every CapEx boom believes its allocation is rational.
Wall Street’s temporary permission to invest is itself a risk. Harry described a $20 billion-plus software company celebrating a 100-basis-point reduction in LLM costs because it lacked permission to sacrifice 5%-10% of gross margin for the category’s best agent: “You’re trapped in a death spiral.” David agreed that positive-return AI functionality should currently trump cost optimization.
4. Meta’s $145 billion plan still lacks a spreadsheetable payoff
Meta reported $56 billion of revenue and $10.44 of EPS versus $6.67, yet the stock was punished after CapEx guidance rose to $125-$145 billion. Google can tie its spend to rapidly growing cloud revenue; Meta’s returns remain indirect.
Meta says model-driven ad improvements are producing perhaps 10%-15% lift, but Rui wanted the missing A/B test. Even if AI adds $10 billion or $15 billion of value, that does not clearly justify spending near $150 billion when some enabling technology might be purchased externally.
The more credible strategic explanation is optionality: if users shift from talking with humans and consuming news toward interacting with chatbots, Facebook intends to be present. That makes the expenditure “a $150 billion bet on the future that’s not quite articulated,” rather than a forecastable new business.
Jason observed that Wall Street can model GPU depreciation and revenue conversion for the hyperscalers, but Meta does not fit those spreadsheets. Rui’s extension: Mark Zuckerberg “doesn’t give a shit about those spreadsheets”; he will spend to remain relevant, with the possibility of another Instagram or WhatsApp—and another failure.
5. Amazon is positioned for the application boom while Apple sits out the frenzy
Forced to buy one and sell one among Amazon, Alphabet, Meta, and Microsoft, Rui weakly chose Amazon over Microsoft. Amazon delivered $181 billion of revenue, AWS reached $37 billion with its fastest growth in 15 quarters, and closer Anthropic alignment gives its distribution engine another source of acceleration.
Jason’s broader bull case was the derivative “application boom.” AI is not merely enhancing existing software; it is enabling more software to be built than at any prior point in their careers, benefiting AWS, GCP, and parts of Microsoft even if token economics eventually mature. Meta does not participate in that infrastructure demand.
The panel nearly skipped Apple because it simply beat across the board, returned capital, and delivered Tim Cook a strong quarter without a developed AI story. Amid the spending frenzy, Apple’s message was effectively: “Thank you, everyone else for getting caught in hysteria.”
Memory inflation complicates every CapEx comparison because some budget increases purchase the same physical capacity at higher prices. The cited consumer specimen was Apple moving the Mac Mini from $599 to $799, though the speaker was uncertain how much reflected the memory configuration; similar costs may eventually reach iPhone pricing.
6. Palantir converted AI urgency into enterprise-scale checks
Palantir’s RPO rose 134% to $4.45 billion, and its Rule of 40 score reached 145%—a level the panel said had otherwise been matched only by AI infrastructure names Nvidia, Micron, and SK Hynix. Alex Karp also spoke of doubling, although Rui could not reconcile his timing with formal forward guidance.
Jason’s mechanism starts with corporate budgeting: every Fortune 500 CEO has “do something in AI” among the top two board priorities, but a $200,000 Harvey deployment or $2 million Sierra project cannot embody enterprise-wide transformation. Palantir can credibly move in $10 million, $20 million, or $100 million increments.
Its government and defense record lets a CEO buy a three-year program, report measurable progress to the board, and close the initiative. Jason contrasted the scale of old enterprise deployments—he cited roughly $26 million for Salesforce and $24 million for Workday at Adobe-scale companies—with Palantir’s ability to deploy transformational work inside a year.
Karp’s striking observation was buying-cycle compression: where one stakeholder once introduced Palantir and others took years to persuade, “every stakeholder shows up to the meeting” now, including the CEO and CFO. Europe’s commercial adoption had barely begun, adding another prospective growth leg.
7. The AI expertise gap is repricing both consultants and operators
Jason called this “the worst gap between in-house and external expertise in our lifetimes.” That scarcity explains why Palantir benefits for years and why consulting arms at Anthropic and OpenAI are not as goofy as they first appear: customers have money and urgency but almost nobody capable of execution.
Jason’s microeconomic defense of consultants was pragmatic: let an expert spend six months learning something and then buy one week of that expertise instead of repeating the work internally. This becomes a Darwin test for agencies—AI-capable HubSpot and Shopify shops may see infinite demand, while undifferentiated peers disappear.
The labor market will bifurcate similarly. Jason sees little use for a $60,000 SDR performing shallow research, but a small class of AI-native SDRs productive enough to replace 20 conventional people could earn $250,000: “Those are the skills you have to have.”
8. Atlassian and Twilio bounced for different—and unequal—reasons
The headline moves were Atlassian up 29%, Twilio up 20%, and Five9 up 23%, but Jason rejected the idea that all three proved a general recovery. Five9’s reacceleration to 9% growth was “not interested in” territory; the two older platforms supplied the real signal.
Atlassian successfully monetized Rovo into its installed base, lifting AI usage and revenue, yet net-new customer growth continued slowing. It passed the first test—customers will pay for an AI product—but may be deferring the harder question of whether the platform can attract incremental demand.
Twilio showed the more interesting second prong. Its disclosures were less precise, but Jason estimated customer count may have expanded roughly 40% over the year as companies such as ElevenLabs and other agent builders adopted its communications APIs, even though ACV did not rise as quickly.
AI builders use Twilio’s APIs, and David said Sierra’s $15 million deal runs on Twilio. Reliability and accumulated carrier infrastructure made the mature product “good enough to benefit from all the trend happening.”
9. SaaS winners need relevance, retention, and preferably new demand
David’s bounded-range view was that mature SaaS companies need not go to zero. When a sound, cash-generative company trades near three times revenue, a credible return toward six times can create a 2x; Atlassian’s 32% GAAP revenue growth supports that bounded recovery, not a new 10x category.
Jason’s stricter two-pronged standard requires monetizing AI among existing customers and attracting new ones. The clearest beneficiaries remain close to infrastructure—Cloudflare, Twilio, Datadog, even DigitalOcean—because new builders consume their products automatically.
HubSpot is the coming test. Its stated plan is to place agents at parity with humans and open the platform fully to them; Jason believes genuine execution should produce reacceleration within 12 months. “It’s a little late, but it’s not too late.” Failure there would darken the outlook for the whole application-software cohort.
David George’s pushback was that mature categories may have exhausted new-logo demand just as Zoom exhausted people needing accounts. Winning can instead mean returning toward 30% growth, selling new products, producing free cash flow, controlling stock compensation, and proving nonzero terminal value—criteria that will also expose the next Medallia-like failure.
10. Anthropic’s revenue ceiling depends on token spend per engineer
Harry placed Anthropic’s revenue pace around $44 billion, or roughly $100 million per day, prompting Jason’s central diligence question: in a mature AI-first company, what percentage of engineering salary dollars becomes steady-state token spend?
Jason’s rough dividing line was 20%-30% versus 5%. At the higher level, Anthropic might grow into several hundred billion dollars of revenue, perhaps even half a trillion; at 5%, today’s scale becomes much harder to reconcile. Coding is the “tip of the spear” because current automation potential is much higher than Jason’s sub-10% estimate for most other knowledge-work functions.
Higher productivity need not reduce developer employment. If $40,000 of tokens doubles a $200,000 engineer’s output, the ROI on engineering rises; a growing company with ten engineers might hire 15, while spending proportionally less elsewhere. “If you can attract them, the number will go up.”
David then complicated the bull case: two semi-autonomous SaaStr agents covering marketing and customer success cost only $254 for an entire month, including $94.27 for the marketing agent. The marketing agent generated better ideas than human peers, though not better execution, suggesting many non-coding workflows will consume far fewer tokens than investors assume.
11. A $50 billion private raise made Anthropic’s IPO optional
Anthropic was reportedly seeking $50 billion at roughly a $900 billion valuation. Chamath admitted his prior call—to skip another private financing and proceed directly to an IPO—was “stupid old-world thinking” when an email can produce commitments within 48 hours without public-market disclosure or liability.
The capital requirement is unusually nonlinear. Chamath’s heuristic was that every dollar of Anthropic revenue requires Anthropic or its partners to invest three or four dollars in compute; forecasting 10x growth a year ahead can mean “committing $30 billion in CapEx for every $1 billion in revenue you have.”
Hyperscalers absorb some exposure, but the model still requires enormous financial guesswork. Chamath’s revised conclusion was categorical: “There is no such thing as too much cash on your balance sheet.” The raise reduces funding risk and gives Anthropic flexibility before an IPO.
The financing probably reduces the odds of a listing this year without necessarily changing its eventual price. Anthropic can still go public if markets, predictability, and preparation align, but it no longer has to; Harry raised the possibility that both Anthropic and OpenAI could slip to 2027, and Chamath agreed that neither now has to go public this year.
12. Sierra proves software still has value—and prices in a huge leap
Sierra was seeking $950 million at a $15.8 billion valuation on approximately $150 million of ARR, or roughly 105x revenue. David George’s concern was that the cited $400 billion customer-service market largely reflects labor, while the existing support-software market may be only $20-$30 billion.
The required thesis is therefore more than replacing Service Cloud. Sierra must capture labor displacement and expand materially into sales and upsell, yet once several AI vendors pursue the same outcome, their economic competitor becomes each other rather than human labor. David believed AI may have expanded the TAM 50%, but was unconvinced it had expanded 10x.
Rory nevertheless saw the round as evidence against “LLMs eat all software.” His estimate was that Sierra’s model costs are probably below 10% of revenue; most value comes from the application, workflow, integration, and domain-specific layer surrounding the LLM. Paying 100x still requires “a very aggressive future.”
Asked where one incremental dollar had more upside, Rory answered “Anthropic” without hesitation: a $6 trillion Anthropic was more plausible to him than a $100 billion Sierra, though he considered both outcomes unlikely. Perceived downside protection around a marquee Sierra asset may help investors psychologically, but the panel warned that venture investors routinely overstate such protection.
13. Musk v. Altman may turn on deadlines and standing, not spectacle
Week one supplied the expected spectacle: Elon Musk conceded xAI had partly distilled OpenAI models, ranked OpenAI and Anthropic above his own under oath, and exposed Greg Brockman’s private diary and estimated $30 billion stake. Roelof found Sam Altman’s zero equity position more surprising than Brockman becoming extremely wealthy as a founder.
The decisive issues may be procedural. Early Musk threats could show he knew enough to sue before the statute of limitations expired, while donations routed through his donor-advised fund may mean the fund—not Musk personally—was the entity legally harmed and therefore the party with standing.
The jury is advisory, leaving the final decision with the judge. Embarrassing testimony may dominate the “tech version of TMZ,” but Roelof thought Musk probably moved backward on the legal merits because those technical defenses can dispose of the case without resolving the larger moral narrative.
14. Coinbase’s “build or go” model raises the bar for every manager
Coinbase’s roughly 14%-15% reduction was framed as removing layers whose members cannot also contribute individually. Jason translated Brian Armstrong’s position as: “If you can’t ship and manage, if you can’t deliver a campaign and be the head of marketing, I don’t want you at Coinbase.”
Roelof normally treats CEOs blaming AI for layoffs as “guilty until proven innocent,” because the explanation often hides overhiring or slower growth. Armstrong earned more credibility from his earlier willingness to keep workplace politics outside Coinbase despite severe backlash; Roelof saw a consistent clarity about culture rather than a fashionable excuse.
Jason demanded more than executives spending 10% of their time experimenting. His autonomous customer-success agent contacted roughly 120 SaaStr sponsors after midnight, captured their problems, and prescribed next actions; a traditional executive might first schedule a meeting that takes two weeks. Likewise, “a CMO today should be able to run their own campaigns” through agents.
The argument ended with a Darwinian prediction: perhaps only 10% of executives can operate this way now, but superior results could make that 20%, then 40%. The same logic informed Harry’s view that work-from-home Friday is often a three-day-weekend device—employees may rationally choose that life, while investors rationally avoid companies organized around it.