Author: Ben Thompson Publication: Stratechery Date: April 13, 2026 URL: https://stratechery.com/2026/mythos-muse-and-the-opportunity-cost-of-compute/
An April 13, 2026 Stratechery essay by Ben Thompson arguing that Aggregation Theory still applies in the AI era, refined to treat the cost of AI compute as an opportunity cost rather than a marginal cost. Thompson frames the recent frontier-lab releases from Anthropic, OpenAI, and Meta as three different bets about what scarce compute should buy.
Summary of argument
Thompson responds to Doug O'Laughlin (Fabricated Knowledge, January 2025), who argued that reasoning models killed Aggregation Theory because compute had become a real marginal cost. Thompson's rebuttal is that the relevant cost is opportunity cost, not marginal cost: AI chips will be used at full utilization regardless, so the strategic question is which workload gets the compute. He casts that as an Aggregation-Theory question, on the view that controlling demand still gives power over supply.
Thompson uses Microsoft's Q2 2026 earnings as his empirical anchor. He notes that CFO Amy Hood described missing the Azure growth expectation not for lack of demand but because Microsoft chose to allocate GPUs to first-party Copilot workloads, which Thompson characterizes as higher gross margin and higher LTV, rather than to third-party Azure customers.
Thompson's conclusion is that Aggregation Theory continues to apply, with refinement: opportunity-cost-of-compute is, in his framing, the new dimension along which aggregator power expresses itself. He argues that the hyperscalers (Microsoft, Amazon, Google) face an internal allocation problem between first-party, third-party, and strategic-investee workloads that they did not face in the pre-AI era.
Three release-strategy comparisons
Thompson evaluates three frontier-lab releases as bets about what scarce compute should buy:
- Anthropic's Mythos — a capability bet on safety and cybersecurity (Project Glasswing), with compute spent on building the most capable model and the safety primitives around it. Thompson rates Anthropic's aggregation position as weak, with no consumer-app moat, and describes Anthropic's bet as relying on enterprise and API revenue plus safety credibility to sustain the model.
- OpenAI's GPT-5.5 — a consumer-experience bet (ChatGPT Images 2.0; persistent multi-tool autonomy), with compute spent on consumer-facing features and the broadest-possible user reach. Thompson rates OpenAI's aggregation position as the strongest among frontier labs, calling ChatGPT a Super-Aggregator candidate.
- Meta's Muse Spark — an engagement-optimization bet, with compute spent on ad optimization and feed personalization. Thompson notes that Meta is already a Super-Aggregator and describes Muse Spark as incremental rather than transformative for Meta.
Key claims
Thompson advances four supporting claims:
- Aggregation Theory's deeper foundation is zero marginal cost, which Thompson argues continues to hold for AI inference, on the grounds that chip electricity is a fraction of chip-amortization cost and that chips run at full utilization regardless.
- The hyperscaler-allocation problem is a new strategic dimension: Microsoft balancing Azure customers against M365 Copilot; Amazon balancing AWS, e-commerce, and its Anthropic and OpenAI strategic investments; and Google balancing GCP, its Anthropic strategic investment, and consumer products.
- OpenAI's strategic position is, in Thompson's reading, structurally weaker than Microsoft's, because OpenAI must keep paying Microsoft for compute even as Microsoft prioritizes its own first-party workloads.
- Anthropic's exit from the Microsoft-OpenAI orbit and into Google and AWS strategic-investment is, by his account, a hedge against this same allocation problem.
Relationships
- depends-on: Ben Thompson
- supports: Aggregation Theory — the AI-era refinement
- related: Claude Mythos Preview
- related: GPT-5.5 ('Spud')
- related: Muse Spark (Meta Superintelligence Labs)
- related: Anthropic
- related: OpenAI
- related: Microsoft
- related: Meta AI
- contradicts: Doug O'Laughlin's "Aggregation Theory is over" framing