Is the “Strongest Model” a Bubble? Winning Benchmarks, Losing Markets
Opinion | November 3, 2026
Author: David Tang, Managing Director, Standard Kepler
Originally published in HKET (香港經濟日報)
Executive Summary
As Anthropic approaches its IPO, AI bubble theories are intensifying. Bull and bear cases both have merit, but two core questions must be answered first: How do we measure the value AI creates? And who ultimately captures that value? Microsoft’s CEO Satya Nadella’s October 2026 observation that “most token usage will trend toward commoditization” provides a critical analytical lens. This report examines Anthropic’s cost structure, the implications of token commoditization, and why the companies best positioned to capture AI value may not be the model makers at all.
The Bubble Debate: Two Competing Narratives
| Position | Argument |
|---|---|
| Bubble Bears | Unsustainable capital expenditure and high leverage. High valuations require profit support; when expectations are disappointed, the bubble bursts. |
| Growth Bulls | AI is transforming every aspect of life. Industry valuations have no ceiling; current spending will generate exponential returns. |
Both perspectives are partially correct. The resolution lies not in choosing sides but in understanding value distribution.
Token Commoditization: The Great Leveling
On October 26, 2026, Satya Nadella stated on the 20VC podcast that most token usage will trend toward commoditization. This does not directly answer whether AI is a bubble, but it reframes the analysis.
What Token Commoditization Means
LLMs handling routine tasks will converge in quality. For general users, large language models will become highly homogeneous. The price differential between models will collapse (currently, Anthropic Fable 5.1 and MiniMax M3 differ by 30-40x in pricing).
The business world does not require the strongest model for every task. A company does not demand that a junior clerk hold a PhD. The vast majority of token consumption will come from cheap, similar-performing LLMs, including open-source models. Only a minority of high-difficulty tasks will require expensive frontier models.
Token commoditization reveals the essence of large model companies’ pricing power — or lack thereof.
Anthropic’s Cost Structure: The Numbers
Anthropic’s IPO filing provides a window into the economics of frontier AI:
2025 Financials
| Metric | Value |
|---|---|
| Revenue | $4.585 billion |
| Compute & Infrastructure | $7.326 billion |
| Compute as % of Revenue | 159.8% |
| Meaning | For every $1 of revenue, Anthropic spent ~$1.60 on compute alone |
H1 2026 Financials
| Metric | Value |
|---|---|
| Revenue | $16.286 billion |
| Compute & Infrastructure | $8.945 billion |
| Compute as % of Revenue | 54.9% |
The improvement is significant but the structural challenge remains: unlike traditional SaaS, marginal costs do not approach zero.
Why AI Economics Differ from SaaS
| Dimension | Traditional SaaS | Large Model Companies |
|---|---|---|
| Development cost | High upfront, near-zero marginal | High upfront, significant marginal |
| Scaling economics | Revenue growth rapidly dilutes R&D | Revenue growth accompanied by ongoing inference costs |
| Profitability path | Clear path to 80%+ gross margins | Margins depend on model efficiency and compute costs |
| Key leverage | User growth | User growth + model efficiency + compute cost decline |
The Litigation Overhang
Beyond compute, litigation and compliance costs are material. Anthropic recognized $1.5 billion in Bartz settlement expenses in 2025. Training large models inevitably encounters copyright and data usage risks.
The critical question: Are these litigation costs one-time events, or will they become persistent profit erosion?
Measuring AI’s True Value: The GDP Connection
Nadella links AI’s long-term value to global GDP growth. Historical parallels provide context:
| Era | Period | Productivity Impact |
|---|---|---|
| Industrial Revolution | ~1760-1840 | UK per-capita GDP growth rose from ~0.2% to peak ~1.25% |
| Information Revolution | 1995-2005 | US labor productivity growth ~2.17% annually |
| AI Revolution | 2023-present | ? |
If AI is genuinely a new productivity revolution, its value will not be captured solely in model company revenues. It will manifest as enterprises producing more value with the same inputs.
The Value Transmission Chain
Compute Investment → Model Capability Improvement → Enterprise Cost Reduction & Efficiency → Per-Capita GDP Growth
AI investment上半场 (first half): The capital expenditure story — opportunities concentrated upstream in chips, cloud, power, raw materials, and data centers.
AI investment下半场 (second half): Downstream transmission — the critical question is how enterprises convert increasingly cheap intelligence into productivity.
The Real Winners: Data Analytics Companies
Companies that transform AI-enhanced intelligence into actionable business value are particularly well-positioned. Their moats derive not from “having lots of data” but from three conditions:
- Data is hard to replicate and continuously updated
- Processing requires domain expertise and proprietary analytical models
- Analysis results enjoy client and market trust
| Company | Ticker | Domain | AI Advantage |
|---|---|---|---|
| TransUnion | TRU | Consumer credit & identity | AI enables deeper risk modeling at scale |
| Equifax | EFX | Credit data & employment verification | Automated income and employment verification |
| Verisk | VRSK | Insurance risk & claims | AI-enhanced catastrophe modeling and claims prediction |
| RELX | RELX | Risk management & legal | Proprietary data embedded in professional workflows |
| Moody’s | MCO | Credit ratings | Brand trust + AI efficiency = broader client coverage |
Conclusion: If There Is a Bubble, What Assumption Needs Correction?
The “strongest model captures most value” assumption may be the bubble. If AI becomes a necessity, tokens may become like electricity: indispensable yet highly homogeneous. Demand will increase, but profits will not rise proportionally.
We should seek the companies that use electricity best, not the largest power plants.
Anthropic’s IPO will undoubtedly become a critical pricing anchor for the AI sector. The phrase “catastrophic or existential risks to humanity” in its prospectus carries unmistakable technological flair. Under intense market attention, its stock may receive initial support.
But as Nadella observed: tokens will commoditize. The correction, if it comes, will not be to the thesis that “AI will change the world.” It will be to the thesis that frontier model makers will capture most of that change.
Standard Kepler Research | standardkepler.com