Anthropic: "our capex is not a bubble" — CIO Krishna Rao joins the debate

Anthropic: "our capex is not a bubble" — CFO Krishna Rao enters the debate

🔎 Why this signal changes everything

The AI bubble debate has just reached a new level. Krishna Rao, the first CFO in Anthropic's history, has come forward with an argument few executives dare to make: the billions swallowed up in GPUs don't come from market euphoria, but from the company's free cash flow, supplemented by debt. No opaque SPVs, no accounting circularity. Real cash flows. The statement lands just as Anthropic prepares its first IPO roadshows — targeted valuation: nearly $1 trillion. If the numbers hold up, it's the first tangible proof that a cutting-edge lab can fund its growth without relying on speculative venture capital. If they don't, the correction will be brutal.

Key takeaways

• Self-funded capex: Krishna Rao claims Anthropic finances its infrastructure through free cash flow + debt, not through speculative fundraising rounds.
• Revenue ×8 since 2024: AI revenue has reportedly grown eightfold in two years, according to the CFO's statements.
• IPO in sight: First investor roadshows held in August 2026, with no valuation discussion at this stage (CNBC).
• Numbers to verify: The IPO prospectus (Reuters, Sept 2026) cites $4.59B in FY2025 revenue; private trackers (GetLatka) cite $65B ARR as of August 2026 — a massive gap to explain.
• Massive debt: $108B raised in three rounds (Sept 2025, Feb 2026, May 2026) + debt = a capital structure that warrants an audit.

Recommended Tools

Indicative prices recorded in October 2026: API and hosting rates change quickly — check up-to-date prices on the official websites before committing.

The CFO Comes Out of the Woods: What Rao Is Really Saying

Krishna Rao is not a communicator. He is an Airbnb veteran who managed the hypergrowth of a marketplace before the IPO. His arrival at Anthropic in 2024 coincided with a valuation of around $20B (PitchBook, cited by WSJ). Two years later, he is leading the first IPO roadshows. His message is simple: "Our capex follows real demand. We are not building for a hypothetical future."

The nuance changes everything. In the dominant narrative — shared by Sequoia, Goldman Sachs, and MIT — AI labs are burning cash to train models whose monetization remains uncertain. Rao flips the argument: it is enterprise customers who drive the capex. API contracts, on-site deployments, and cloud agreements (AWS, Google Cloud) generate cash before the GPUs are ordered. The cycle is inverted: revenue → capex, not capex → hope → revenue.

But the claim of being "funded by free cash flow" deserves scrutiny. Anthropic raised $108B in fresh capital across three giant rounds between September 2025 and May 2026, at valuations climbing from $183B to $965B. Debt is added on top. If free cash flow covers part of the capex, debt and equity cover the rest. The question is not binary: what exact proportion? The IPO prospectus (Reuters, Sept 28, 2026) will have to clarify that.

The "×8 since 2024" figure: reality or run rate?

Rao claims an eightfold multiplication of AI revenue since 2024. Public sources give divergent benchmarks:

The quadrupling in six months (Dec 2024 → May 2025) is documented by Reuters. Rao's "×8 since 2024" could be comparing early 2024 (~$200M according to GetLatka) to mid-2026 (~$47-65B ARR). But comparing unaudited private ARR to GAAP revenue is misleading. ARR counts signed contracts, not yet invoiced. Run rate annualizes the current month. GAAP revenue is recognized according to accounting standards. Three metrics, three stories.

VentureBeat explicitly reminds us of this: private run rate "is neither audited GAAP revenue, nor gross margin, nor free cash flow". Rao, as a good CFO, knows the difference. His essay — if it exists — must specify which metric he is using. Without that precision, the "×8" remains a marketing claim, not an accounting fact.

The capex structure: debt, equity, and the circularity trap

The core of Rao's argument: no circularity. No "Microsoft invests in OpenAI, which buys Azure, which inflates Microsoft's revenue." Anthropic has no proprietary cloud. Its models run on AWS (strategic partnership, $4B invested by Amazon) and Google Cloud. API revenue comes from third-party customers — enterprises, startups, developers. No closed loop.

Yet debt changes the equation. Raising debt to buy GPUs assumes that future cash flows will cover debt service. If revenue falls — enterprise slowdown, API margin compression, technological disruption (e.g., high-performing open models) — the debt becomes a ball and chain. The $108B raised in equity provides a cushion, but dilutes founders and employees. At a $965B valuation, each billion raised now represents only 0.1% of the capital. It's efficient... as long as the multiple holds.

The IPO prospectus (Reuters, Sept 2026) mentions "sharply rising costs." Normal: training Claude Opus 4.7 (a score of 90 on LLM General, 94.3 on Agentic) costs tens of millions per run. Large-scale inference costs even more. The gross margin on the API is not public. If it's low — as with many AI SaaS companies — free cash flow stays negative despite growing revenue. Rao will have to show the margin curve, not just the revenue curve.

The context: why this debate now?

Three forces are converging. First, public markets are growing impatient. The "Magnificent Seven" have carried the S&P 500; their combined AI capex exceeds $200 billion in 2026. Investors are demanding visible ROI. Second, open-source competition is intensifying. DeepSeek V4 Pro (score 88), Kimi K2.6 (84), and GLM-5.1 (83) offer free or low-cost alternatives. Third, enterprises are slowing down: frozen budgets, POCs that never convert, "proof of concept" fatigue. Our analysis of Microsoft reveals the true cost of AI documents this adoption plateau despite billions invested, and the case of Tesla, which caps its employees' AI spending at $200/week, shows that even tech giants are now rationing access to AI tools.

Anthropic is not alone in this. Google DeepMind bled dry: Nobel laureate John Jumper joins Anthropic, Transformer architect Noam Shazeer heads to OpenAI — the AI talent war enters a brutal phase illustrates the pressure on teams. Anthropic accuses Alibaba of the largest distillation attack ever documented: 28.8 million exchanges, 25,000 fraudulent accounts — the AI war enters a new dimension shows that intellectual property is under threat. Anthropic Dreaming: Claude agents learn from their dreams between sessions proves that product innovation continues. But the economics must follow.

What the IPO prospectus will have to reveal

An S-1 (or F-1 for a Cayman Islands company like Anthropic) forces transparency. Here are the line items Rao cannot avoid:

1. GAAP revenue by segment: API, enterprise, cloud partner (AWS/GCP), other. No aggregated ARR.
2. Gross margin by segment: Inference cost vs. billed price. Including GPU depreciation.
3. Detailed free cash flow: Maintenance capex vs. growth capex. Debt service broken out.
4. Customer concentration: Top 10 customers = what % of revenue? Churn risk.
5. Capex commitments: Contracts signed with NVIDIA, AWS, GCP for 2026-2028. Firm amounts vs. options.
6. Stock-based compensation: Actual annual dilution. Impact on shareholder FCF.
7. Related-party transactions: Amazon, Google, strategic investors. Transfer pricing.

Without this data, "funded by free cash flow" remains an assertion. With this data, the market will judge. The August 2026 roadshows (CNBC) are the test: institutional investors will ask these questions face-to-face. Their reaction — engagement or withdrawal — will be worth more than any essay.

The "No Bubble" Trap: Surviving Your Own Success

Saying "no bubble" when you're worth $965B is risky. If the market corrects — say -40% on AI stocks — Anthropic finds itself at ~$580B. Employees whose RSUs are underwater leave. Recruiting seizes up. Google DeepMind Bled Dry shows that talent follows traction, not theoretical valuation.

Rao knows it. His job isn't to inflate the multiple, but to build a company that survives the correction. That means: gross margin > 60%, positive FCF before IPO, revenue diversification (not 80% API), a durable technical moat. Diversification is becoming urgent: even the defense market is slipping out of Anthropic's monopoly now that Trump is imposing multi-vendor AI at the Pentagon. Anthropic Dreaming suggests R&D that widens the moat. But R&D is expensive. The virtuous cycle (revenue → R&D → better model → more revenue) only holds if margins allow it.

What the Competitors Don't Say

OpenAI doesn't publish FCF. xAI (Grok 4.1, score 90 General / 79 Agentic) is burning Elon's cash. Google (Gemini 3.1 Pro, 92 General / 87.3 Agentic) buries AI in its Cloud results. Meta (Llama, not in the rankings) open-sources to commoditize the model layer. Microsoft, for its part, is playing both sides: historic OpenAI partner while deploying its own frontier models, as exemplified by MAI-Thinking presented at Build 2026. Only Anthropic, through its CFO, is attempting the "we are a real business" narrative. That's bold. It's also the only path to a credible IPO in 2026-2027.

But beware: the "real business" narrative attracts regulators. FTC, DOJ, EU, UK CMA are scrutinizing cloud partnerships (Amazon-Anthropic, Microsoft-OpenAI). If capex is "funded by free cash flow," but that free cash flow comes from cloud contracts with strategic investors at non-market prices, circularity reappears in legal form. The prospectus will have to disclose the AWS/GCP terms. Total transparency is the only shield.

❌ Common Mistakes

Mistake 1: Confusing ARR, run rate, and GAAP revenue

What's wrong: Citing "$65B in revenue" (GetLatka, August 2026) as an accounting fact.
The fix: Always specify the metric. ARR = annualized signed contracts. Run rate = last month × 12. GAAP revenue = recognized under accounting standards. The three tell different stories. Use the IPO prospectus ($4.59B FY2025) as the audited anchor.

Mistake 2: Taking "funded by free cash flow" at face value

What's wrong: Repeating the claim without asking for the capex / FCF / debt breakdown.
The fix: Demand the detailed cash flow statement. Maintenance capex (GPU replacement) vs growth (new clusters). Debt service. SBC. Shareholder FCF = FCF - SBC - debt service. That's the number that matters.

Mistake 3: Ignoring ongoing dilution

What's wrong: Looking at the $965B valuation without calculating the % held by founders/employees after 3 giant rounds + IPO.
The fix: Model the dilution. $108B raised at ever-increasing valuations = massive dilution. If founders hold < 10% post-IPO, the alignment changes. The prospectus will show it.

Mistake 4: Believing that the absence of a proprietary cloud eliminates circularity

What's wrong: Saying "no circularity because no cloud".
The fix: Check the AWS/GCP commercial terms. Cloud credits converted into revenue? GPU transfer pricing? Minimum purchase commitments? Circularity can be contractual, not just structural.

❓ Frequently Asked Questions

Is Anthropic really profitable?

No, not in the GAAP net income sense. The CFO talks about positive free cash flow to fund capex, not net profit. API gross margin is not public. The $4.59B in FY2025 revenue (prospectus) likely masks massive operating losses (R&D, S&M, GPU). Positive FCF would be a first for a frontier lab — to be verified in the S-1.

Why raise $65B in May 2026 if FCF covers capex?

Because FCF covers part of capex, not all of it. Debt and equity fill the gap. Above all, raising at a $965B valuation costs little in dilution (0.1% per billion). It's prudent cash management: building a war chest before the IPO, repaying expensive debt, funding international expansion — latest example, Anthropic's opening in Seoul and the MOU signed with South Korea. Not contradictory with Rao's narrative — but nuanced.

Is the "×8 since 2024" verifiable?

Not with current public sources. The quadrupling from Dec 2024 → May 2025 ($1B → $3B annualized, Reuters) is documented. The ×8 would compare early 2024 (~$200-500M ARR) to mid-2026 ($47-65B ARR GetLatka). But comparing unaudited private ARR to GAAP revenue is methodologically wrong. Wait for the IPO prospectus for a consistent series.

Will the IPO happen in 2026?

Unlikely. The August 2026 roadshows (CNBC) are preliminary — "testing appetite", not "launching the offering". A confidential S-1 filed in late 2026, IPO in H1 2027 is more realistic. Conditions: open markets, demonstrated positive FCF, no geopolitical shock (Taiwan, GPU export controls). Rao is preparing the option, not the date.

What happens if enterprise demand collapses?

Debt becomes critical. If API revenue drops 30%, FCF turns negative. Debt service (coupons, amortization) must be paid. The $108B in equity absorbs losses... until it runs out. At a $965B valuation, the safety margin seems wide. But tech history (Cisco 2000, telecoms 2001, SPACs 2021) shows that 100x revenue valuations can correct 80% in 18 months. Committed capex (firm GPU contracts) then becomes a fixed burden.

✅ Conclusion

Krishna Rao is right on the substance: Anthropic has a real business, real customers, real enterprise traction. But "funded by free cash flow" is an accounting claim still awaiting its audit. The IPO prospectus will settle the matter. Until then, track quarterly shareholder FCF, not marketing ARR. The AI bubble won't deflate on a CFO's say-so — it will deflate when the marginal cost of intelligence falls below the API's selling price. That day is drawing near. Anthropic Dreaming shows that R&D is accelerating. The economy, for its part, always catches up in the end.