Capital Cools, Capability Climbs: AI's Story Splits in Two

2026-11-06·6 min read·OKCodex Editorial

Capital Cools, Capability Climbs: AI's Story Splits in Two

Put this week's signals side by side and you get an uncomfortable conclusion: model capability is still climbing, but the capital and policy curves that fund it are flattening or tightening. This is not a collapse narrative — it is a divergence narrative. The same industry is showing very different slopes depending on which part you look at.

1. Capital: usage is rising, spending is not

Anthropic's IPO has hit resistance, with investors pushing back on the $1.5 trillion valuation while the tech market sold off. On the other side, Ramp data cited by TechCrunch shows AI usage still climbing while spending has stalled — driven by price cuts and tighter cost controls among the biggest spenders.

Together, these point to one problem: usage growth and revenue growth are no longer the same curve. Trading price for adoption is a classic software path, but when unit prices fall faster than usage rises, valuation assumptions built on hypergrowth need recalibration. That does not necessarily mean demand has peaked — it does mean the linear extrapolation from "more usage" to "more revenue" no longer holds.

2. Policy: 41 ballot measures and the local cost of compute

During the US midterms, voters weighed in on 41 data-center-related ballot measures. In all but one, the majority voted to impose more restrictions on data center development. Data centers are the heaviest and least reversible asset class in this buildout. Local permitting friction will not show up in model capability immediately, but it will show up in the long-run marginal cost of compute.

3. Capability: no slowdown, but the price is explicit

Thinking Machines, the startup from former OpenAI CTO Mira Murati, is releasing Inkling 2 Small, a mixture-of-experts model with 594 billion parameters. The company claims it is the strongest western open-weight model. The notable part is the cost: better agentic performance comes with sharply rising token usage. Open weights lower the licensing barrier; they do not automatically lower the inference bill.

A subtler signal comes from METR. The nonprofit spent eleven weeks auditing Anthropic's internal research data and concluded that the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. However, Anthropic redacted the most informative evidence, so the public cannot independently verify it. "Development is accelerating itself" is a significant claim, but without visible evidence it remains something worth watching rather than something established.

4. Applications: always-on agents get a price tag

Perplexity is launching "Teammates," always-on agents that work through ongoing projects on their own virtual machines. The feature requires the Max plan at $200 per month — well above similar offerings from OpenAI or Anthropic — and the always-on agent market is getting crowded. Meta, meanwhile, introduced Spaces as a new way to collaborate with its Muse agent. The direction is consistent: agents moving from one-off chats to persistently online work units.

Always-on means always billing, and always consuming compute. With spending stalled and compute costs exposed to policy, the $200 price is simultaneously a test of willingness to pay and a public statement about cost structure.

Our read

These are not one story. The capability curve is still steep, the capital curve is flattening, and the policy curve is tightening. Bundling them into either "AI bubble" or "AI keeps exploding" is too coarse. What is actually being tested is unit economics — under simultaneous pressure from rising token consumption, falling unit prices, and climbing marginal compute costs, who can push the cost of a single task below the revenue it generates.

What to do about it

  1. Recompute your token budget. If your plan is based on per-task token consumption from six months ago, agentic upgrades have likely invalidated that number. Measure before you price.
  2. Track usage and spending separately. Ramp's data shows they decouple; use two independent metrics for external market reads and internal cost accounting.
  3. Evaluate open-weight models yourself. A third-party "strongest open model" claim is a starting point. Validate on your own task set, and watch actual token spend in agentic scenarios.
  4. Treat compute policy as a long-horizon variable. Local permitting trends will not change tomorrow's bill, but they will shape inference cost assumptions two to three years out.
  5. Hold audit-style conclusions loosely. METR's findings cannot be independently verified because key evidence was redacted. Adjust your attention accordingly; do not treat it as settled fact.

Sources

Share:Weibo

Subscribe to weekly briefing

Comments

Loading comments…

Leave a comment

0 / 500

Comments appear after review.

Find tools for your next task

Start with scanned documents, code review or research. Get a shortlist with matching evidence; no model is run.

Find tools for my task →
Analyze this topic with two models

DeepSeek and Qwen are preselected with the article topic. You choose when to run them.

Try it →

Keep reading

Explore

Explore published agents