GAVIN BAKER: A fragmented, hybrid multi-model AI market can let open-weight models, frontier labs, infrastructure and applications all succeed

The Gist

Gavin Baker says everyone can win: cheap open models handle most of the work, the big labs still grow, and chip and cloud companies benefit from more competition. Shortages keep the boom from turning into a bust. This is a steelman: Gavin Baker's thread is reconstructed in its strongest fair form, not endorsed.

Conclusion

The most likely outcome is a hybrid multi-model market in which open-weight models serve most tokens, leading frontier labs still do well though less dominantly than seemed possible in May, and AI infrastructure and applications gain from model-layer competition and lower margins, while supply constraints guard against an overbuild.

Premises

Fragmentation and competition at the model layer are better for AI infrastructure suppliers and application companies than a two-lab oligopoly, and the leading labs can still grow within a competitive market.
As long as someone funds training, lower margins at the model layer mechanically shift more of each end-user dollar to compute and raise compute demand, so broad open-weight share is the best case for AI infrastructure demand and a competitive field of several frontier labs is the next best.
A hybrid, multi-model enterprise pattern is emerging in which owned specialized models sit beside frontier models, and this pattern lowers prices at the model layer while preserving or raising demand for compute.
The headline $413 billion of Anthropic non-cancellable contracts is large but reasonable relative to its duration and Anthropic's scale, because the cloud portion is about $31 billion a year that operations can fund and the Broadcom portion is being externally financed; the same test should be applied to OpenAI once its filing is public.
Repricing of below-market compute contracts should accelerate cloud and neocloud revenue and ease financing concerns, so external financing as a share of the AI buildout may be near its peak.
Constraints on power, wafers and capital make the AI buildout slower and smoother, which lowers the risk of overbuilding and is better for investors and the public over the long term; a severe cyber event is the more plausible near-term risk.
Disaggregated inference could extend the economic life of Blackwell GPUs well beyond six years, which supports infrastructure return assumptions.

Assumptions

Baker manages Atreides Management, which co-led Fireworks AI's Series D and invests in AI-related companies.
The forecast depends on volume growth outpacing efficiency gains and on continued funding of frontier training.
2026 credit markets showed stress for AI infrastructure borrowers, which tempers the financing part of the forecast.

Analysis

Argument type: Inductive.

Potential Fallacies

Conjunction of contingent claims (Conclusion) — The conclusion bundles open-weight majority share, frontier-lab success, infrastructure and application gains, and no overbuild into a single 'most likely outcome'. Each added element lowers the joint probability, and the argument does not compare this scenario against alternatives such as renewed concentration, commoditization that hurts labs, or a demand-led bust.
Share versus absolute level (Premise 2 (and Premise 3)) — A larger fraction of each end-user dollar going to compute does not by itself mean more total compute spending. That requires demand to be elastic enough, and volume to grow faster than efficiency gains cut compute per token. Calling the effect 'mechanical' treats a conditional empirical outcome as an accounting identity.
Token share versus value share (Conclusion; Premise 3) — 'Open-weight models serve most tokens' is compatible with frontier labs keeping most revenue and margin, or with the model layer's revenue pool shrinking. Without a stated metric, the claim that open weights dominate and the claim that frontier labs do well are not shown to fit together.
Possibility offered as support (Premises 1, 5, 7; Conclusion) — Premises phrased as 'could extend', 'can still grow' or 'may be near its peak' are modest claims, but the conclusion treats them as support for a 'most likely' outcome. A series of possibilities does not add up to a probability.
Desirability versus likelihood (Premises 1-2 → Conclusion) — Premises 1 and 2 rank scenarios by how good they are for infrastructure and applications ('best case', 'next best'). They do not supply a mechanism or evidence that the best case for those layers is also the most probable outcome.
Assuming the contested point (circular affordability test) (Premise 4) — Calling $31 billion a year 'fundable from operations' presupposes the revenue growth and pricing power that are in question, especially given the argument's own prediction that model-layer prices and margins fall. The test is useful but has no downside case or threshold, so it may not discriminate between sound and unsound commitments.
Constraint treated as a pure stabilizer (Premise 6) — Power and wafer limits slow physical deployment, but they do not cap capital committed ahead of demand. Historical buildouts (fiber, for example) overbuilt despite bottlenecks, and constraints can ease abruptly or trigger over-ordering. The claim holds only if demand-side discipline is also present.
Investor interest equated with public interest (Premise 6) — 'Better for investors and the public' assumes the two align without argument. A smoother buildout may leave externalities such as energy costs, security risks and concentration of compute unaddressed.

Counterarguments

These challenges are not ranked.

Premise 2 (High impact) — If open-weight models take most tokens at thin margins, who pays for the next frontier training runs? Open-weight models largely follow frontier research, so defunding the frontier could stall the whole ecosystem and flatten compute demand growth. The scenario depends on a funding condition that its own dynamics undermine, and at the limit, free models would be the 'optimum' under the argument's logic, which is absurd.
Premise 2 (High impact) — Cheaper tokens raise total compute spending only if demand elasticity exceeds one. Distillation, better architectures, quantization and disaggregated inference can cut compute per token faster than volume grows, shrinking the revenue pool for the whole stack.
Conclusion (High impact) — 'Everyone wins' is the least stress-tested outcome. Plausible alternatives include bifurcation (a few labs capturing premium agentic work while the rest commoditize), renewed concentration after a capability jump, or a shakeout in which over-committed labs and weaker neoclouds fail and damage infrastructure suppliers.
Premises 4 and 5 (Medium impact) — The two pull against each other: repricing below-market contracts raises costs for labs, straining the operating cash flow that Premise 4 says funds their commitments. Non-cancellable obligations are fixed while revenue is variable, so a revenue shortfall or credit shock turns a manageable commitment into a fragility.
Premise 6 (Medium impact) — Telecom and fiber, and many commodity cycles, show that supply constraints delay but do not prevent overbuilding, because capital is committed ahead of demand and scarcity encourages over-ordering. Constraints that ease at the same time demand disappoints would remove the protection exactly when it matters.
Premise 7 (Medium impact) — In a power-limited environment, occupying scarce power with older, less efficient GPUs is costly. Rapid product cadence may make older chips uneconomic well before six years, and longer GPU lives also reduce replacement demand for chip vendors and weaken used-GPU price floors that collateralize neocloud debt.
Premise 3 (Medium impact) — Multi-model use may be an early cost-optimization or hedging phase that consolidates later, especially if one lab reaches a decisive capability threshold in agents or long-horizon work. The sentiment swing between May and October itself shows how quickly the consensus can flip.
Conclusion (Medium impact) — Open-weight supply is not guaranteed. It depends on a small set of sponsors (hyperscalers, chipmakers, Chinese labs) whose motives can change with strategy, regulation, export controls or a severe cyber event that disproportionately hits models that cannot be recalled.

Suggested Improvements

Define the market-share metric — State whether 'open-weight serves most tokens' refers to token volume, revenue or profit, over what time horizon and in which segments (enterprise, consumer, API), and give a threshold such as a share by a date. Token share and value share diverge sharply. Without a metric the conclusion can absorb almost any evidence and its coherence with 'frontier labs still do well' cannot be judged.
Address the training-funding loop — Explain who funds frontier and open-weight training as model margins compress, and estimate the margin range in which the scenario holds. This is the weakest link. Resolving it would convert the proviso 'as long as someone funds training' from an escape clause into a testable condition.
Quantify demand elasticity versus efficiency — Provide data or scenarios on price elasticity of inference demand and on compute-per-token trends, so the claim that lower prices raise compute demand is shown rather than assumed. This parameter decides whether the infrastructure thesis holds, and the argument concedes it in its own assumptions.
Stress-test the commitment analysis — Add downside revenue cases, contract schedule escalation and counterparty concentration to the $413 billion test, and set a threshold for what counts as 'reasonable'. Apply it to OpenAI with the same inputs. A test that nearly any long-duration commitment can pass does not discriminate. Downside cases matter most for non-cancellable obligations.
Provide evidence for the enterprise pattern — Cite spend surveys, customer examples or aggregated token data showing hybrid deployments at scale, drawn from outside the author's own portfolio network. 'Emerging' is an anecdotal claim that currently carries most of the market-structure forecast.
Give probabilities and falsifiers — Replace 'most likely' with scenario weights for alternatives (oligopoly, commoditization bust, bifurcation) and name observable indicators that would change the forecast, such as GPU rental prices, credit spreads and lab funding rounds. This would reduce the unfalsifiable feel of a scenario in which almost every layer wins, and align the confidence of the conclusion with its hedged premises.
Distinguish winners within layers and broaden the stakeholders — Acknowledge that a layer can grow while individual firms fail, and address energy, ratepayer and security externalities rather than equating investor and public welfare. Aggregate layer-level success can mask significant dispersion and cost-bearing, which matters for the 'better for investors and the public' claim.

Scenario Tests

Efficiency gains (distillation, better architectures, disaggregated inference) cut compute per token faster than usage grows. (Challenges) — Total compute spending flattens or falls despite a larger share of end-user dollars going to compute. The infrastructure benefit, the GPU life support and the financing outlook all weaken, which is the failure condition the argument itself flags.
Open-weight models capture most token volume while frontier labs retain most revenue through premium, agentic or high-stakes workloads. (Supports) — The 'open-weight serves most tokens' and 'frontier labs still do well' claims are reconciled, but the open-weight claim then means little for value, and infrastructure gains come mostly from volume.
Margin compression at the model layer reduces funding for frontier training and open-weight sponsors stop releasing frontier-class weights. (Challenges) — The capability engine slows, enterprise reasons to adopt newer models weaken, and compute demand growth stalls. The scenario's central condition fails through its own mechanism.
A frontier lab achieves a decisive capability lead in autonomous long-horizon work and enterprises pay a premium for it. (Challenges) — The hybrid pattern re-concentrates toward an oligopoly. Infrastructure still benefits from demand growth, but the fragmentation premise and the 'less dominant' lab claim fail.
Credit stress deepens and Anthropic's revenue growth falls short of its contract schedule. (Challenges) — Non-cancellable commitments become a source of counterparty distress for clouds and Broadcom, contradicting 'financing near its peak' and the reasonableness verdict on the $413 billion.
Power and wafer supply stay tight while demand keeps growing steadily, and compute repricing continues to rise. (Supports) — Constraints do pace the buildout and support infrastructure revenue and financing, though much of the rent may accrue to bottleneck suppliers rather than clouds and neoclouds.
A severe AI-enabled cyber event triggers regulatory restrictions on open-weight distribution. (Challenges) — The author correctly names this as a risk, but it would hit open-weight models hardest and undermine the open-weight-led outcome the conclusion predicts.

Coherence & Relevance

The argument is a loosely linked cluster of three parallel sub-arguments: market structure (hybrid multi-model, open-weight share), economics for infrastructure and applications (margin transfer, repricing, GPU life), and financing and supply (commitment test, constraints). The economics sub-argument is the most coherent, since it offers a conditional mechanism and acknowledges its dependencies. The market-structure sub-argument has the largest gap, because no premise establishes that open-weight models will serve most tokens or that labs will do well rather than merely can. Several premises are in internal tension: falling model margins versus the training funding they require, repricing versus lab affordability, and power scarcity versus extended GPU life. The title's modest 'can let ... succeed' is also weaker than the conclusion's 'most likely outcome'. Overall, the argument is a plausible, informed scenario with useful analytical tools, but it is not a demonstration that this scenario is the most likely one.

Fragmentation and competition at the model layer are better for AI infrastructure suppliers and application companies than a two-lab oligopoly, and the leading labs can still grow within a competitive market. (Strong) — Establishes who benefits from fragmentation, not that fragmentation will occur or that open weights will serve most tokens. 'Can still grow' is a possibility, while the conclusion asserts that labs will do well.
As long as someone funds training, lower margins at the model layer mechanically shift more of each end-user dollar to compute and raise compute demand, so broad open-weight share is the best case for AI infrastructure demand and a competitive field of several frontier labs is the next best. (Strong) — Supplies the core mechanism for the infrastructure benefit, but moves from share of the dollar to absolute demand without an elasticity argument, and ranks scenarios by benefit rather than probability.
A hybrid, multi-model enterprise pattern is emerging in which owned specialized models sit beside frontier models, and this pattern lowers prices at the model layer while preserving or raising demand for compute. (Strong) — The most direct support for the 'hybrid multi-model' label, but it lacks data and does not show that open-weight models serve a majority of tokens.
The headline $413 billion of Anthropic non-cancellable contracts is large but reasonable relative to its duration and Anthropic's scale, because the cloud portion is about $31 billion a year that operations can fund and the Broadcom portion is being externally financed; the same test should be applied to OpenAI once its filing is public. (Moderate) — Bears on financing health and the absence of a lab-driven bust, not on market structure. It depends on revenue trajectories that the argument's own price-compression thesis puts under pressure.
Repricing of below-market compute contracts should accelerate cloud and neocloud revenue and ease financing concerns, so external financing as a share of the AI buildout may be near its peak. (Moderate) — Supports the infrastructure-health sub-claim, but sits in tension with credit stress and with Premise 4, since repricing raises lab costs. The 'near peak' call is a timing forecast with no metric.
Constraints on power, wafers and capital make the AI buildout slower and smoother, which lowers the risk of overbuilding and is better for investors and the public over the long term; a severe cyber event is the more plausible near-term risk. (Strong) — Directly supports the 'supply constraints guard against an overbuild' clause, but does not address demand-side overbuild risk or the historical record of constrained overbuilds, and the cyber ranking is unsupported.
Disaggregated inference could extend the economic life of Blackwell GPUs well beyond six years, which supports infrastructure return assumptions. (Weak) — Peripheral to the market-structure forecast and speculative in itself. It bears only on infrastructure returns, and conflicts somewhat with efficiency gains and power scarcity.

View this argument on LogicFirst.ai