80% of revenue comes from new models: How vulnerable is Anthropic?
The bottom line: The market currently sees Anthropic as the next NVIDIA and Meta, with extremely high growth expectations. If growth falters, its valuation and share price could fall by 25%, 30%, or even half.
In addition, Anthropic’s revenue currently depends heavily on its models’ technological edge. The company still has a large amount of computing capacity scheduled to come online over the next two years, and cost pressures may continue to mount. 2028 to 2029 could mark the real turning point toward profitability. Before then, Anthropic must race ahead and build up a large enough base of enterprise customers, making the large language models’ historical memory and switching costs its moat. Once switching costs for enterprise customers rise, Anthropic can break free from its absolute dependence on having the most powerful models, ultimately supporting its current valuation.
I. Huge valuation and operating reality
Claude’s parent company Anthropic has filed for an IPO and plans to list on Nasdaq under the ticker ANT. The market had previously discussed a $2 trillion valuation, but that has recently been lowered to around $1.5 trillion. Even so, Anthropic could still rank among the world’s top ten by market capitalization at the time of its listing.
At present, the number of IPO shares, pricing, financing size, total market capitalization, and the first-day free float have not been disclosed.
•
If a 25x price-to-sales (P/S) ratio is used to support a $1.5 trillion market cap, annual revenue of $60 billion is required; to support a $2 trillion market cap, $80 billion is needed.
•
Historically, companies that experienced similar explosive growth include Nvidia, Facebook, and Snowflake.
What is Anthropic’s current operating situation?
•
2025: full-year revenue of $4.585 billion, up 1,088% year over year.
•
First half of 2026: revenue of $16.286 billion, up 1,300% year over year, achieving 3.55 times last year’s full-year revenue in just six months.
•
Q2 2026: revenue of $11.555 billion, up 144% quarter over quarter, and the company recorded quarterly operating profit for the first time (operating profit of $974 million, margin of 8.4%).
However, compared with Meta and Nvidia, Anthropic still has a sizable gap. Reuters reported that Anthropic’s third-quarter revenue was about $16.8 billion (up 45% quarter over quarter, annualized at $67.2 billion). In contrast, Meta’s 2025 revenue exceeded $200 billion, and Nvidia’s fiscal 2026 revenue exceeded $215 billion—both about three times Anthropic’s current scale—and both already have mature and stable profitability. By comparison, Anthropic still posted an operating loss of $693 million in the first half of 2026, with compute spending accounting for as much as 54.9% of revenue.
Using Reuters’ third-quarter revenue figure, Anthropic’s trailing twelve-month revenue is about $35.1 billion. At valuations of $1.5 trillion, $1.75 trillion, and $2 trillion, the corresponding price-to-sales ratios are roughly 43x, 50x, and 57x, respectively. For reference, Nvidia’s current P/S ratio is about 17x, Snowflake’s is about 22x, and Palantir’s is about 78x.
If a 25x P/S ratio is used as the long-term normalized value, that means trailing twelve-month revenue would need to grow from $35.1 billion to $60 billion–$80 billion, an increase of 71% to 128%. However, the estimated third-quarter revenue growth has already slowed to 45% quarter over quarter, which is also the main reason institutions have lowered valuation expectations from $2 trillion to $1.5 trillion; this shows how sensitive the market is to growth rates.
II. Core risk: strong dependence on the competitiveness of new models
The prospectus discloses a key figure: models released in the past six months contributed more than 80% of revenue.
•
Positive side: once a new model is recognized by the market, it can rapidly drive customer migration and spending growth.
•
Potential risk: the company must continuously invest heavily to maintain its technological edge. In other words, current profitability is extremely sensitive to model pricing and customer usage.
Under the current cost structure, if core model revenue declines by 10% to 20%, profits could be severely compressed, even pushing the company back into operating losses. The company explicitly notes that model versions are typically updated on a monthly basis; once it falls behind by one or two versions, a clear gap in performance and commercial competitiveness may emerge.
Profit sensitivity stress test (based on Q2 2025)
•
Baseline data: revenue of $11.555 billion, compute costs of $5.558 billion, operating profit of $974 million.
•
Scenario assumptions: 80% of revenue is affected by model competition; 70% of compute costs are fixed in the short term and 30% vary with usage; 80% of sales expenses vary with revenue, while other fixed costs such as personnel remain unchanged.
•
A 10% price cut: quarterly operating profit would plunge from $974 million to $186 million (a decline of about 81%).
•
A 20% price cut: even if usage stays unchanged, quarterly operating profit would directly turn into a $600 million loss.
How can the impact of price cuts be offset?
To respond to a 20% price-cut scenario:
1.
Compute costs need to fall by 10.8% for the company to return to break-even.
2.
Compute costs need to fall by 28.3% to restore the original $974 million profit.
However, improvements in computational efficiency do not mean actual costs can decline immediately, because much of the compute capacity is already locked into long-term contracts. Even if the compute required for each task decreases, the company may still need to pay the original fees. The freed-up compute resources must generate new revenue or truly reduce cash outlays in order to improve financial performance.
Anthropic clearly recognizes this as well, and is gradually reducing its dependence on expensive short-term cloud compute through direct chip leasing and building dedicated compute capacity:
•
In April 2026, the company announced an expansion of its TPU partnership with Google and Broadcom, with additional capacity expected to come online starting in 2027.
•
By 2028 to 2029, as new compute clusters have run for several quarters, utilization and scheduling efficiency improve, and inference technology advances, unit compute costs are expected to decline. Anthropic previously projected internally that the company would achieve positive cash flow in 2028.
III. Moat and long-term investment logic
Can Anthropic survive the compute expansion phase and achieve stable profitability? If it can get through the first few years, economies of scale may mean Anthropic will no longer be overly dependent on the single label of being the “strongest model.”
As Claude Code, Cowork, and enterprise-grade agents continue to penetrate enterprise customers, the ability to reliably complete real work tasks becomes crucial. At that point, memory and switching costs will build a true moat.
As of the third quarter:
•
Over the past twelve months, the number of enterprise customers spending more than $100,000 has reached 11,900.
•
Of these, more than 1,900 spend over $100 trillion.
•
About 250 spend over $1,000 trillion.
•
Among customers with annual spending above $100,000, more than half also use Claude Platform or at least one Claude-owned application (up from about 30% at the end of 2025).
•
Net revenue retention (NRR) is about 900% (note: because the statistical starting point only requires customers to generate at least $1 in revenue, this more reflects the rapid expansion of spending among early customers; the long-term stickiness of mature customers still needs further validation).
IV. Conclusion: cheers and concerns across the semiconductor supply chain
Before reading the prospectus, the market may have been somewhat pessimistic, because the current market structure resembles an inverted pyramid—the foundation at the bottom used to be Nvidia, but now it has been replaced by Anthropic and the not-yet-listed OpenAI.
However, along with the prospectus release, semiconductor stocks such as Nvidia, Broadcom, AMD, and Micron generally rose, because Anthropic publicly disclosed massive compute procurement plans for the next few years:
•
Broadcom-related TPU equipment leasing: total payments over five years of about $161.2 billion.
•
AMD-related compute deployment: expected spending of more than $20 billion.
•
In addition, Anthropic has also procured substantial Nvidia GPU compute from multiple cloud providers.
For the semiconductor industry, 2027 may be a year when Anthropic’s own profit margins come under pressure, but it will also be a year when semiconductor orders are recognized in a concentrated manner.
However, if Claude’s revenue growth slows in the future and utilization of newly added compute is insufficient, Anthropic will inevitably reduce the next round of purchases. Therefore, future earnings reports—especially revenue performance, customer usage, utilization of purchased compute, and whether the next round of compute spending continues to expand—will directly determine whether upstream semiconductor companies can further lift their valuations, and whether these massive compute demands can truly translate into revenue, EPS, and earnings estimate upgrades for upstream firms.