The AI market isn't collapsing. It's digesting

Stop conflating a product kerfuffle with infrastructure cycles and scaling dynamics

There’s a tidy story making the rounds: GPT-5 underwhelmed, therefore, the AI bubble is popping. It’s convenient, and wrong. The argument splices together four different layers of reality and treats them as one: (1) product experience choices; (2) app-layer ROI; (3) infrastructure capex and supply chains; and (4) scientific progress and scaling. When you disaggregate the stack, what looks like a collapse resolves into a digestion phase with predictable pains and equally predictable carry-through.

Below is a clean framework to reason about it.

Layer 1: Product experience ≠ capability frontier

Most of the heat around GPT-5 is about how it feels: tone shifts, refusal regimes, and routing that sometimes lands you on a safer-but-duller persona. Those are product decisions (alignment knobs, guardrails, model routing, defaults), not scientific regressions. The paradox of safety-tuning is that modest capability gains can be perceived as losses if the UI shifts the “voice” or clamps down on unsafe edge cases.

Three common confusions:

• Alignment vs ability. If the model avoids certain outputs or shortens chains of reasoning in risky domains, users infer “worse model.” You’re measuring a policy, not the underlying aptitude.

• Routing vs. core model. Assistants often mix models. If the router misfires to a smaller/safer model for some requests, users experience variance that they attribute to the flagship.

• Expectation inflation. A released framed as a “new epoch” sets a reference point that real-world deltas rarely match. The delta still matters, especially in coding, longer-horizon tasks, and tool use, even if it’s less cinematic than the hype reel.

Verdict at this layer: a product-management controversy, not evidence of a capability stall.

Layer 2: App-layer ROI is messy, and orthogonal to infra demand

Yes, a ton of enterprise pilots under-deliver. Reasons are mundane, not metaphysical: data plumbing isn’t ready; workflows aren’t redesigned; incentives fight adoption; risk controls slow deployment; wrapper apps are commoditized. This is the application shake-out we should expect when a general compute primitive floods the market before process re-engineering catches up.

Important distinctions:

• Wrappers vs systems. Wrappers (chat-in-front-of-API) commoditize quickly. Systems that rewire workflows (agents, tool use, retrieval over proprietary data, knowledge graphs) win more slowly but more durably.

• Horizontal vs vertical. Horizontal copilots have marketing reach; vertical copilots with deep domains tools have ROI. The latter force you to co-design with operators and accept longer sales cycles; they’re also where moats form.

• Metrics that matter. Minutes saved and customer satisfcaction scores are vanity metrics without P&L hooks. Track error-adjusted throughput, working capital turns, or incident rate deltas. If you can’t tie to dollars, you won’t survive the next budget committee meeting.

This layer can look bleak while the infra layer hums along: weak PMF at the edge doesn’t nullify the need for training cycles, model upgrades, or agentic scaffolding. It simply reallocates who captures value.

Layer 3: Infra capex is shifting from sprint to digestion, not off a cliff

The infra story is not “spend stops.” It’s “mix and utilization matter more from here.” Three simultaneous truths:

1. Capacity still ramps. Power, HBM, and advanced packaging (CoWoS-class) remain gating factors. Even with aggressive expansions, many regions stay tight through 2026. That props up pricing and encourages continued buildouts.

2. Utilization risk is local, not systemic. Hyperscalers with full-stack demand (search, ads, cloud, office, consumer apps) can keep GPUs busy as they rotate models and soak up test-time compute. The wobble risk concentrates in independent AI clouds and specialized providers exposed to lease escalators, concentrated customers, and short-dated offtake. They can hit an air pocket even while aggregate demand rises.

3. From sprint rent to term structure. As the market matures, expect less spot-renting of flops and more structured offtake (longer contracts, minimum volume commitments, indexed pricing). Financialization of compute—futures/forwards, capacity reservations, hedges—smooths digestion and lowers the cost of capital for both builders and buyers.

“Capex digestion” means: rate-card compression where supply outpaced near-term customers; stronger emphasis on power procurement, interconnects, thermal and availability SLAs; and a premium on utilization management. That’s a far cry from collapse.

Layer 4: “Scale is all you need” was never a serious thesis

The caricature says: “We scaled pretraining; we’re done.” The serious program now looks like scale * algorithmic ideas * test-time compute * tool use * better data. The frontier is shifting:

• Test-time compute and planning. More chains of thought, external memory, verifiers, and search/planning loops. Gains here don’t require bigger base models so much as smarter computation at inference.

• Tool use and agents. Calling code, databases, and services, not as a demo, but as the default runtime, turns models from text predictors into action systems.

• Data quality and curricula. Curation, synthetic data regimes, RL from outcome metrics, and curricular tuned to tasks move the needle when naive scaling saturates.

Diminishing returns to naive pretraining does not imply the end of progress. It shifts what you scale.

If you fuse the four layers: a redesign of assistant defaults spawned a “vibes” backlash; many app startups are hitting the pilot-to-value wall; infra spend is rotating and professionalizing; research progress is pivoting from unstructured scale to structured reasoning. None of that is a bubble popping. It’s the shape of a power-law technology maturing.

What would actually falsify the “digestion” thesis?

Keep a short, falsifiable checklist:

1. Two+ consecutive capex downgrades from multiple hyperscalers, not just mix shifts, with explicit deferrals of data-center builds rather than resequencing.

2. Sustained under-utilization measurable in public rate cards (price wars on top-tier GPU clusters) alongside idle capacity disclosures or obvious GPU holidays.

3. Plateau in reasoning-heavy benchmarks and agentic workflows across successive releases and a stall in tool-use reliability (e.g., code-execution win rates).

4. Power procurement freeze: interconnect queues and substation delays translate into widespread cancellations rather than schedule slips; FERC/utility constraints hard-cap buildouts.

5. Enterprise renewal cliff: two or more cohorts where AI suite/copilot renewals materially underperform initial pilots after serious workflow integration attempts.

If 2–3 of these trigger together, reassess. Short of that, expect rotation rather than rupture.

Strategy notes

For allocators

• Favor bottlenecks over logos. Memory (HBM), substrate, advanced packaging, grid upgrades, transformers/substations, and thermal. These remain rate-limiting for longer than people think.

• Scrutinize independents’ capital stacks. Watch lease escalators vs. revenue clauses, customer concentration, and contract tenor. A 90-day revenue tail on 36-month liabilities is how you recreate WeWork in silicon.

• Prefer term-linked exposure. Capacity reservations and compute offtakes with step-downs, floors/ceilings, and power passthroughs. This is where underwriting improves fast.

For founders (app layer)

• Build systems, not wrappers. Agentic workflows wired into proprietary data and domain tools. Measure dollars, not demos.

• Own evaluation. Evals are product, not research. If you can’t prove outcome deltas on your customers’ metrics, you’re a feature.

• Exploit test-time compute. Shipping a small base model with a strong planning/search scaffold can beat brute force on cost and quality.

For infra builders

• Secure power first. Land is easy; megawatts aren’t. Substation lead times and interconnect queues kill schedules more reliably than chip slips.

• Design for utilization. Multi-tenancy isolation, topology awareness, and job schedulers tuned to agentic workloads will differentiate. “90%+ sustained utilization” is a sales asset now.

• Hedge the cycle. Don’t pretend volatility is a surprise. Lock term contracts where you can; keep optionality where you can’t.

Why the collapse narrative feels persuasive, and why it isn’t

• Narrative myopia. A single flagship model’s vibe change is legible to everyone; HBM supply curves are not. We overweight the visible.

• Hype hangover. Expectations compound faster than science. When the release isn’t science fiction, we update too far in the other direction.

• Category conflation. We bundle “my copilot didn’t change my life” with “no one will buy GPUs.” Different markets, different clocks.

The digestion metaphor helps: the industry gorged on capacity, prototypes, and narratives. Now it has to absorb and reorganize. That looks like slower headlines, tougher procurement, more rigorous ROI conversations, and a step-function improvement in how we use models rather than how much compute we can pour into pretraining. It also looks like capital structures becoming sane: less spot, more term; less vibes, more underwriting.

Call it a comedown if you need to. Practically, it’s the phase where moats are dug, weak balance sheets crack, and durable operating leverage shows up in the boring places: power, packaging, interconnects, and the unglamorous craft of making agents actually do work.

Bottom line: If your thesis hinges on “GPT-5 disappointed me, therefore the bubble bursts,” you’re diagnosing vibes, not markets. Separate the layers, watch the falsifiers, and plan for digestion, not detonation.

---

If you enjoy this newsletter, consider sharing it with a colleague.

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)