---
title: "Anthropic: Can't escape 'intelligence deflation' — endgame: AI Google or AT&T?"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/dolphin/post/45175577.md"
description: "IPO pricing looks reasonable. But 'cognitive deflation' is compressing the future profit pool."
datetime: "2026-11-19T08:30:18.000Z"
locales:
  - [en](https://longbridge.com/en/dolphin/post/45175577.md)
  - [zh-CN](https://longbridge.com/zh-CN/dolphin/post/45175577.md)
  - [zh-HK](https://longbridge.com/zh-HK/dolphin/post/45175577.md)
author: "[Dolphin Research](https://longbridge.com/en/dolphin.md)"
generator: "portal-rs"
---

# Anthropic: Can't escape 'intelligence deflation' — endgame: AI Google or AT&T?

In the prior piece '[Anthropic: Shedding the cash burner to become a mortgage slave](https://longbridge.cn/topics/44562870?channel=SH000001&invite-code=355628&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=b6dbea64-0b7b-4657-b274-fb8b4e50c35f)', Dolphin Research ran the numbers on $Anthropic(ANT.US). Nearly $1 tn of compute commitments are now sitting on the balance sheet, and revenue needs to rise 4–5x to carry the load. Once the math is done, two more fundamental questions emerge.

First, what sustains the 'intelligence premium'. Token pricing has dropped about 70% in a year, and open-source models are closing fast from behind. Can Anthropic’s premium hold up.

Second, where is the revenue ceiling. Primary markets marked Anthropic at $965 bn in May, and IPO chatter peaked near $1.5 tn. At that price, what sort of 2028 are buyers underwriting.

This piece centers on Anthropic’s valuation case. It addresses growth prospects and moats for Anthropic and AI more broadly through the two questions above. The article has four parts.

**1) Where do bulls and bears really diverge on Anthropic’s valuation.**

**2) How big is the AI model TAM, and how to think about brain-power deflation.**

**3) Will frontier models commoditize, and how long can Anthropic’s moat hold.**

**4) How to value the business.**

**Detailed analysis**

**I. Valuation dispute: what is the core divergence.**

$1.26–1.41 tn is the fully diluted valuation implied by the IPO range, or roughly 16–17x Oct ARR of $81 bn. Using Barclays’ small-group readout last week and annualizing Q4 revenue to ~$77 bn, the multiple is closer to 18x. Layer on a sequential downtrend and 18x P/ARR is hardly cheap.

The decline is mainly attributed by the company to holiday seasonality. That explanation is not unreasonable, but for a lab leader touting trillions in incremental revenue headroom, invoking seasonality at this stage feels odd. How do you justify a high-growth multiple with that.

Yet on the roadshow, investor demand does not look as cautious as feared. Indications reportedly exceeded the deal size by 6x. Anthropic quickly became a poster child for bulls and bears calling each other idiots.

**Dolphin Research reviewed views from public and private investors and found the core split sits in how TAM is framed for Anthropic.** **a) Bull case: up to $3 tn valuation, TAM anchored to $40 tn in wages.**

Investors such as Gavin Baker and Brad Gerstner define TAM similar to Anthropic’s S-1, using global knowledge-worker wages as the reachable pool. Labor income is ~52% of global GDP, or about $60 tn. Anthropic demarcates knowledge roles (from coding to general cognitive work) and argues roughly $40 tn of wages are in scope for AI-led transformation.

On that framing, $100 bn-scale revenue is a small bite of a $40 tn reachable TAM. A $3 tn valuation can still look cheap. The logic rests on broad salary pools rather than near-term spend.

**b) Bear case: down to $500 bn valuation, TAM anchored to $6 tn in enterprise IT budgets.**

Skeptics tie Anthropic’s target market to enterprise IT budgets. Under the bear view, Anthropic captures 3% of total IT spend five years out, or ~$180 bn in revenue. At 5x mature software P/S and discounting at a 15% IRR, that maps back to roughly $400–500 bn today.

**But the future path for AI is too uncertain for long-dated anchors to settle the debate.** Most investors default to a medium-term revenue build and a not-excessive multiple, yielding a neutral $1.0–1.5 tn range. That aligns with IPO pricing, and the latest fully diluted read is near the top, at $1.5 tn.

**II. TAM: mapping to today’s markets risks a fatal distortion.**

Recent US restrictions on remote compute could slow China’s open-source model iteration by limiting high-end chips. That may bolster closed models like Anthropic and ease token price pressure for a while. Even so, from a forward supply-demand lens, sustained token price declines look inevitable.

**It is not just competition. The technology stack itself pushes in the same direction.** **a) Chips double every generation.**

With the same 1 GW of power, newer chips yield more tokens. Hardware cycles lift throughput, lowering unit costs with each node. That advantage compounds as deployments scale.

**b) Model algorithms keep improving efficiency.** Anthropic’s filing notes algorithmic gains in training and inference that let the same compute produce more effective output. Each new model is trained with help from prior models, and stronger models train stronger successors while also delivering cheaper service.

This recursive self-improvement (RSI) widens gaps at the frontier in both capability and cost-performance. It reinforces scale and raises the bar for laggards. The cycle favors incumbents with access to capital and data.

\<a-b> capture dual iterations on supply and demand that drive token prices down 45–55% per year, implying 10–15% QoQ. Over the past year, Opus input prices fell from $15 to $5, then $4 per mn tokens, down ~73%. Third-party indices show Anthropic’s transaction prices halved in a month from $1.70 to $0.83.

New export rules could temporarily slow the decline during strict enforcement, possibly compressing Avg. QoQ cuts to sub-10%. The trend, however, remains downward. Price cuts will test volume elasticity.

Sustaining revenue growth hinges on whether price declines unlock enough incremental demand to overcompensate. In early phases, lower prices reduce adoption barriers and accelerate diffusion. This is the Jevons paradox bulls often cite when AI growth is questioned after each model upgrade and token cut.

**Dolphin Research argues demand elasticity varies by use case and industry.** **(1) Expansionary (elasticity > 1): cheaper drives more use and higher total spend.**

**Coding is the archetype.** Building an internal tool used to take two months and two devs, so many requests sat idle. Now an agent can ship in an afternoon, surfacing long-tail demands: rewrite legacy code, tool up every workflow, audit codebases end-to-end.

These were cost-suppressed demands, and cheaper tokens unlock multiples of usage. Beyond accelerating IT penetration in traditional industries, frontier science and finance also fit this bucket. Innovation cycles compress as friction falls.

**(2) Adversarial (elasticity ≈ 1): spend tracks rivals and sector dynamics.** **Examples include litigation, cybersecurity, and marketing.**

Competition temporarily inflates demand, but as ROI erodes and industries consolidate, spend plateaus. If opposing counsel uses AI to crank 10 briefs a day, you must match. Hackers use AI to find exploits at scale, forcing AI-native defense.

Here, budgets are set by what opponents spend and what the task is worth, not by unit price. Lower prices buy more service within a fixed budget and invite new entrants, compressing margins. As the sector profits shrink, involution-driven demand recedes.

**(3) Deflationary (elasticity < 1): volume fixed, price cuts just reduce client spend.** **Think audit, tax, compliance, payroll.**

A listed company does one annual audit and quarterly tax filings. They will not file more because it is cheaper. KYC is one check per client, and cheaper contract review does not make firms sign 10x more contracts.

Savings flow to P&L or are redeployed elsewhere. The pie does not expand in those workflows. Unit economics accrue mostly to buyers.

**As Anthropic’s growth logic leans on substituting knowledge work, two questions persist for skeptics.** **a) Most roles are not expansionary.**

Within the $40 tn knowledge worker market, expansionary use cases are not the majority. Using US workforce and departmental budget distributions as a proxy, roughly 64% of headcount and 56% of wages sit outside IT and frontier science. Those areas are less likely to expand total spend.

**b) Under brain-power deflation, even expansionary cases turn deflationary faster.** Coding is expansionary today because traditional firms have low IT penetration and budgets can rise naturally. But AI’s cognitive edge is orders of magnitude. At current rates, a developer’s salary equates to 113 Claude Team premium seats.

Output is not yet fully substitutive, but time savings on general tasks and cycle times on long tasks are already material. A year ago AI could handle about two hours of coding; now it is at two working days, doubling roughly every four months. About one in five new PRs in public repos now comes from Claude Code.

**Back-of-envelope: one AI can replace 113 × 1/5 = 22.6 humans, saving ~20x in IT wage costs.** With such deflation in cognitive value, the market can expand on penetration over the short to medium term. Over the long run, the $40 tn brain market shrinks even before full penetration.

**Thus Anthropic’s S-1 growth narrative, whether anchored to the $40 tn wage pool or to <1% penetration of Fortune 500 IT budgets, misses a critical bug.** **The cognitive value of current human labor will deflate materially over time as AI substitutes.**

**III. Competition: value migration depends on whether Anthropic becomes AT&T or Google.**

**Token price-for-volume mirrors telco ‘faster and cheaper’ dynamics.** Over a decade, mobile data prices fell 90%+, and per-capita usage rose by tens of times. Yet carrier mobile revenue stayed roughly flat, implying elasticity near 1.

Where did enterprise savings go. The internet’s value map shows it flowed to Google and Meta, which sit atop the traffic stack. These applications own 1P traffic and the channel.

Once network services commoditize with low switching costs, winners must invest in channels to reach users and in ecosystems to retain them. Touchpoints and content keep attention. AT&T had user reach but missed the application layer and never became a Google or Meta.

If AI model APIs analogize to telco networks, will the industry evolve similarly. **Dolphin Research’s view.** Anthropic’s technical edge shows most in long-task completion, where compounded step accuracy is critical: a 1% per-step error rate almost guarantees failure over 500 steps.

Long-task success depends on context capability (effective memory) and high-quality data (task trajectory data with outcome signals). Anthropic adds a further closed loop. By Sep 2026, over 90% of its codebase commits came from Claude, with each generation helping build the next.

**Anthropic aims to maintain and widen gaps via supervised partial self-iteration.** **How fast that flywheel spins is constrained in the short run by available compute.** Supply determines iteration cadence.

**(1) Context capability.** It is not about who has the longest context window (frontier models cluster around 100M), but who recalls most accurately and curates what to store. ‘Remember longer and recall better’ gaps have narrowed.

In an early-2026 test retrieving eight items from 1 mn tokens, Claude led GPT and Gemini by ~76% vs ~30%. By H2, GPT-5.6 and Meta’s Muse reached 70–98%, and Opus 5.5 stopped disclosing this metric. Claude’s edge now lies in context management engineering and the vast real task trajectories in Claude Code.

These are engineering capabilities in the tool layer, not the model’s native memory. The distinction matters for portability and defensibility. The moat sits above the base model.

**(2) High-quality data.** That means complete real-task trajectories plus RL environments that can auto-judge and correct. Claude Code’s scale provides a steady feed of such data.

**Claude’s lead concentrates in long coding and AI R&D tasks that are long-chain and verifiable.** In METR tests, Opus 4.6 completes ~12 expert hours independently, about 2x GPT and Gemini. In Vals AI’s latest run, Opus 5.5 matches Gemini 4 overall but leads by 5–10 pts on hard terminal tasks, autonomous AI R&D, and a 5-day space mission, at roughly a quarter of rivals’ cost.

The lead is not universal. On a year-long business simulation requiring commercial judgment, Gemini and GPT come out ahead. This underscores that Claude’s advantage stems from coding scenarios with abundant traceable trajectories and verifiable rewards, defining the moat’s boundary.

That moat may only last 1–2 generations. Open-source lags closed by ~4–7 months, about one to two closed-model release cycles. Once a frontier model ships, it can be distilled by fast followers.

**Since H2, Anthropic and OpenAI have tightened anti-distillation measures.** Beyond hidden chain-of-thought, they now ban accounts, block replayed reasoning, and deploy classifiers to intercept, all to shrink supply of high-quality trajectory data. Anthropic claims a 90%+ drop in siphoned data by Oct.

This shows up in release cadence. Open-source monthly gains (ECI per month) slowed from 1.8 to 0.8, while closed-source accelerated from 1.7 to 2.7, widening the gap from 5.6 to 12.8 pts. The divergence is concentrated in long tasks.

On public benchmarks and short tasks, open-source still reaches 97–98% of closed. But on KSP long-task tests, open-source is ~11% of closed performance. **The root cause is RSI speed: using models to train models compounds advantages.**

Laggards can still catch up using open high-quality datasets plus internal labeling. But iteration slows materially, and if RSI remains effective, closing the gap gets harder over time. So the outlook bifurcates by use case.

a) General-purpose models for copy, support, translation, and simple coding likely drift toward ‘infrastructure’. They will bill per unit, and prices will fall annually. b) Frontier models will not commoditize quickly.

Frontier edge depends on closed-loop self-iteration and sits where demand also values the edge. Complex coding, investing, and legal are adversarial or expansionary scenarios where clients pay up. But that edge requires tens of billions per year in training capex.

Anthropic expects training spend to exceed $90 bn by 2027. Fall behind for one generation and you drop into the generic tier that competes on price. The treadmill is unforgiving.

**(3) Closed-loop iteration sustains advantage.** Continued lead derives from internal high-frequency use and self-iteration. This is a ‘use more, get stronger’ learning loop that presumes ample compute.

Anthropic has been aggressive in pre-buying compute for 2025–2026. While it did not tout a ‘Stargate’ like OpenAI, it has been busy since H2 2025. Disclosed contracts and chip orders imply ~11 GW in future capacity, above OpenAI’s 10 GW Stargate.

**2\. From model to product and ecosystem is the real moat.**

AT&T’s lesson: reach does not equal entry. Without terminal applications and an ecosystem, a pure API becomes the AI era’s ‘pipe’. Software-defined services have low switching costs.

a) Clients call Claude via Cursor, Copilot, or various SaaS, and user intent is expressed in someone else’s UI. b) Swapping models can mean changing one line of config. MCP-like open standards and listings across all three major clouds lower switching costs to ‘number portability’ levels.

**Durable moat comes from using technical lead to build a business ecosystem.** It shows up in three areas. **1) Entry:** users state ‘what to do’ directly in Claude Code and the Claude app, akin to Google’s high-value intent box.

**2) Accrual:** long-term memory, project context, team skills, and workflows that settle in Claude are sticky. The longer you use it, the harder it is to leave, like Meta’s social graph. **3) Channel:** as OS, Office, and browsers underpinned Microsoft, Apple, and Google, Anthropic must become the default in enterprise workflows.

It needs 1P traffic as a channel, not a component embedded in others’ products. To get there, Anthropic must accept competing with some early customers. Claude Code will have to compete directly with API customers like Cursor.

AT&T once protected SMS revenue by shying away from internet apps. If Anthropic protects API revenue and avoids building the entry, it risks the same fate. Long-term value depends on whether it becomes AT&T (pipes) or Google (ecosystem).

**Its terminal valuation between $1 tn and $3 tn will hinge less on tokens sold and more on where user intent is expressed and whether the ecosystem closes the loop.** Expect a battle with end-user channels over the entry point.

**IV. How to judge Anthropic’s value.**

Sections II and III already frame the valuation levers: TAM and share. But the focus shifts by horizon. **a) Near to mid term (1–2 yrs): supply.**

Compute is tight near term. Anthropic has pre-booked out to year three; spot compute costs run $40–50 bn per GW annually, close to self-build on a per-year basis. Rent is annualized opex, while self-build spreads $50 bn over 5+ years.

Hence near-term TAM is compute-limited. On share, in an early market with first-mover technical advantage (self-learning iteration), Anthropic’s position should hold. Share depends on signed capacity and inference allocation.

**b) Long term (5 yrs): position in the stack.** Once compute is abundant and costs fall, revenue is driven by position in the value chain. That determines reachable TAM and share.

If Anthropic is ‘Google-like’, it taps the ad pool equivalent of its ecosystem. If it is ‘AT&T-like’, it earns a cost-plus on traffic. We model a neutral short term and two long-term cases: optimistic (Google) and pessimistic (AT&T).

**1\. Neutral short-term case.**

The New York Times reports Anthropic at 3.5 GW in compute by year-end, doubling to 7 GW by next year-end. By 2028, Living with Agents cites a total 11.3 GW. Dolphin Research, based on contracted capacity and disclosed 1P/3P GW, estimates 5.1 and 9.3 GW in 2027 and 2028.

The difference is end-year vs period Avg.: media cite end-year points, while our contract/term/$ per GW math yields period Avg. capacity. Averaging end-year figures also aligns closely with our estimates. Assuming token price cuts of ~70% for new models over the past year and a slower 50% cut in 2027 due to RASA limits, plus chip/model efficiency doubling per 1–2 years, revenue per GW still trends down.

Under these assumptions, 2028 revenue lands below $200 bn. On valuation, we use 8–10x P/S, or ~50–60% of growth SaaS at 15–20x. Discount reflects a heavier model.

Anthropic must build capacity and the balance sheet to stay ahead and keep funding training. Unlike mature SaaS with 80%+ GPM and 50–60% OPM at scale, Anthropic’s profitability is structurally lower. At 8–10x P/S, the 2028 terminal value is $1.6–2.0 tn.

Discounting at 15% to end-2027 implies a neutral range of $1.3–1.6 tn. At $1.0 tn+ the IPO looked very attractive; at $1.5 tn now, upside for long-only shrinks, and risk-reward above $2.0 tn worsens. Discipline matters around exuberance.

**2\. Long-term optimistic case.**

This assumes Anthropic leverages model advantage to expand its ecosystem. It shifts from ‘selling APIs (compute+model)’ to ‘selling an ecosystem’ that spans chips, compute, and channels. It ends up one of three leaders at the AI application layer.

Winning the ecosystem requires five conditions, which we benchmark against rivals. The snapshot highlights Anthropic’s biggest gap. **It lacks a dominant terminal entry to end users.**

Beyond weaker control of AI infrastructure, it has no phone, no OS, no billion-user app, and as a B2B player lacks CSP-scale channels. Its likely share of the ‘three’ is the enterprise and developer agent entry alongside Google and OpenAI/Microsoft.

The optimistic path: Claude Code evolves from a developer tool to Cowork, a workspace spanning sales, legal, and finance. Then it becomes the default platform for enterprise-deployed agents, with MCP as the de facto interface standard. At that point, Anthropic charges ecosystem rents, not just tokens.

Using the earlier expansionary/adversarial/deflationary split, assume 25% AI work adoption and AI costs at 1/10 of labor. If usage scales 2x, 0.5x, and flat across the three buckets respectively, AI’s share of enterprise costs totals ~$1.14 tn. If Anthropic takes one-third, that implies ~$380 bn revenue run-rate.

At 10x P/S, the terminal value is ~$3.8 tn. Choice of discount rate and horizon depends on risk appetite. This embeds some brain-power deflation.

**3\. Long-term pessimistic case: stuck at the model layer as an ‘infrastructure’ commodity.**

Here, Anthropic fails to build ecosystem advantage during its lead. As model progress slows and rivals catch up, the sector gets stuck at the model layer. AI models commoditize into a cost baseline.

**Prices keep falling until a thin margin over compute costs remains.** Today Anthropic’s 3P leased compute costs roughly $14 bn per GW per year. If revenue per GW year gravitates toward $14 bn, expanding capacity on pure model sales ceases to make sense.

Absent an ecosystem pivot, the likely outcomes are a sale to a player that owns the entry or to an app-layer leader. Assume model-layer OPM converges to 10%, akin to mature industrials or regulated utilities. Anthropic’s signed compute contracts imply peak compute cost of $172.2 bn in 2029 on straight-line amortization.

Add other opex at 15% of revenue and back-solve revenue to ~$230 bn in 2029. If sold then, at a 20% tax rate, net income is ~$18.4 bn. At 10x P/E (commodity or mature telco/industrial), equity value is ~$184 bn.

Add net cash north of $100 bn to total ~$300 bn. **This is an extreme case that assumes rapid homogenization and value collapse within three years.** Today Anthropic likely has at least a one-year absolute lead.

**Stricter US controls on upstream AI inputs also extend Anthropic’s lead window.** Unlike AT&T’s missteps, Anthropic is already building out its ecosystem. It still lacks a super-scale end-user platform, but the posture is different.

**Ultimately, avoiding an operator fate requires building the enterprise ‘work entry’.** Employees should open Claude Code or Cowork first, and enterprises should default to Claude and MCP for agent deployment. Unlike phones dominated by OS, enterprise entry is more contestable.

Anthropic has recently launched multiple Office-like products to that end. It is pushing for default status in enterprise workflows. That is the path to durable value capture.

**Final thoughts.**

**Dolphin Research’s neutral 2028 valuation is $1.6–2.0 tn (discounted at 15% to end-2027/2026 gives medians of ~$1.6 tn/~$1.4 tn). The optimistic terminal is ~$3.8 tn, and the pessimistic terminal is ~$0.3 tn.** Given current operations, a $1.5 tn IPO is reasonable for the 1–3 yr horizon.

But low earnings visibility is a key bug. Analyses that lean on multi-trillion or tens-of-trillions TAMs often ignore the risk of brain-power deflation. That does not preclude near-term scarcity premia for a rare listed asset or upper-tail outcomes under uncertain endgames.

Like SpaceX, valuation can trade into the optimistic zone in hotter markets. Even in our optimistic case with some deflation effects, crossing $3 tn requires Anthropic to secure a spot among the top three at the app layer. It is premature to underwrite that today.

For secondary investors, excess returns from long holds will come from correctly judging the end-state ecosystem slot. Recent compute export limits temporarily hit open-source, which could support stronger post-IPO sentiment for longer than expected. That tailwind may buy time.

**Lastly, Anthropic’s unlock schedule resembles SpaceX’s staggered approach to mitigate overhang.** The biggest shocks tended to be the first and the largest unlocks, especially when paired with exuberant post-listing rallies. Markets become more ‘immune’ to later unlock waves.

For this IPO, the base primary issue is 960 mn new shares (ex co-founder secondary) at 83 yuan. Strong demand triggered a full green-shoe in new shares and a concurrent private placement, raising nearly $120 bn. About 15.5 bn shares are locked, roughly 12x the initial float, with 25 scheduled unlock windows over one year.

**The three largest waves fall in early Mar (the second trading day after FY26 results), mid-May (Q1 print or day 180, whichever earlier), and Nov (Q3 print or day 365, whichever earlier).** Watch those windows and manage trading risk around them. Stay nimble into catalysts.

\<End>

**Dolphin Research’s prior Anthropic coverage:**

Nov 11, 2026 sell-side roadshow take '[Anthropic mini-conf.: lots of key info, has guidance management begun.](https://longbridge.com/zh-CN/dolphin/post/44975810?channel=OWNN00210)' Oct 27, 2026 initiation (Part I) '[Anthropic: From cash burner to mortgage slave, the trillion-dollar march is long](http://longbridge.com/zh-CN/dolphin/post/44562870?channel=OWNN00210)'

Risk disclosure and disclaimer: [Dolphin Research Disclaimer and General Disclosure](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### Related Stocks

- [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md)
- [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md)
- [T.US](https://longbridge.com/en/quote/T.US.md)
- [ANT.US](https://longbridge.com/en/quote/ANT.US.md)
- [GOOGM.US](https://longbridge.com/en/quote/GOOGM.US.md)
- [GOOGN.US](https://longbridge.com/en/quote/GOOGN.US.md)
- [T-A.US](https://longbridge.com/en/quote/T-A.US.md)
- [T-C.US](https://longbridge.com/en/quote/T-C.US.md)
- [GGLL.US](https://longbridge.com/en/quote/GGLL.US.md)
- [GGLS.US](https://longbridge.com/en/quote/GGLS.US.md)

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**