Agents Over Bubbles (Stratechery Article 3-16-2026)

Ben Thompson | March 16, 2026

---

1. Key Themes

Theme 1: The Three LLM Paradigm Shifts — and Why the Third Changes Everything

Thompson maps three inflection points: ChatGPT (capable but unreliable), o1 reasoning models (reliable but still human-directed), and now agentic systems (autonomous, self-verifying, and human-optional in the loop). Each paradigm multiplied compute demand, but the third does so non-linearly.

"It's the third paradigm, however, that has truly tipped the scales in favor of capex expenditure not being speculative investment but rather badly needed investment in meeting demand that far exceeds supply. First, generating an answer will often entail multiple calls to a reasoning model. Second, the agent itself needs compute... Third, agents are another step function increase in usefulness, which means they are going to be used even more than even reasoning models in a chatbot."

---

Theme 2: Agents Collapse the "Adoption Bottleneck" — Demand Scales Without Mass Human Buy-In

The traditional concern about AI adoption has been that most consumers won't proactively use it. Thompson argues agents invert this dynamic: a small number of highly agentic users can direct enormous compute workloads, making mass adoption irrelevant to demand growth.

"What this means in terms of compute — and by extension, economic impact — is that it actually won't require that many people with agency to drastically increase the amount of compute that is actively utilized to create products with meaningful economic impact... the rise of agents doesn't just mean a dramatic increase in compute, but also a narrowing of the need for widescale adoption by humans for that demand to manifest."

---

Theme 3: Model + Harness Integration Is the New Profit Moat — Commoditization Thesis Is Wrong

Thompson argues that the "models are commodities" thesis, plausible in Paradigm 1, is invalidated by the agentic era. The harness (agent orchestration layer) must be tightly integrated with the model to deliver results. This integration is where differentiation and profit pool concentrate — analogous to Apple's hardware/software integration.

"Model performance isn't the only thing that matters: the integration between model and harness is where true agent differentiation is found... profits flow away from modular parts of the value chain — which are commoditized — and flow towards integrated parts of the value chain, which are differentiated."

---

Theme 4: Enterprise Is the Real AI Revenue Engine — and Agents Turbocharge the Case

Thompson argues enterprises — not consumers — are the natural buyers of agentic AI, both because they pay for productivity and because agents directly address the coordination-cost drag of large organizations.

"Anthropic got it right by focusing almost entirely on the enterprise market: companies have a demonstrated willingness to pay for software that makes their employees more productive, and AI certainly fits the bill... What makes enterprise executives truly salivate, however, is the prospect of AI not simply eliminating jobs, but doing so precisely because that makes the company as a whole more productive."

---

Theme 5: The Coming AI-Driven Workforce Contraction Is Structural, Not Cyclical

Thompson warns that the coming wave of layoffs attributed to AI won't just be post-COVID normalization. Agents move the optimal "headcount equilibrium" point permanently downward, and the companies that cut most aggressively will have structurally lower cost bases.

"More and more companies are not simply going to wonder if they hired too much for a pre-AI world, but also if they hired too much for a post-AI world; the most forward-looking and future-proof approach will likely be to cut more rather than less, with the hope that those who remain have no choice but to rebuild scale with agents. After all, if they don't, dramatically smaller competitors built with AI from the beginning will soon be nipping at their heels."

---

2. Contrarian Perspectives

Contrarian 1: "We Are Not in a Bubble" — And the Capex Is Justified, Not Speculative

The consensus hedged view has been that AI spending may reflect bubble dynamics. Thompson breaks from his own prior framing (he previously argued "bubbles can be good") to assert there is no bubble at all. His evidence: every hyperscaler reports demand exceeding supply, and the structural drivers (agents, enterprise imperatives, workforce replacement) provide fundamental demand justification.

"Sitting here in March 2026, however... I've come to a different conclusion: I don't think we're in a bubble... the economic returns from using agents aren't just impactful on the bottom line, but the top line as well. In this context, is it any wonder that every single hyperscaler says that demand for compute exceeds supply?"

---

Contrarian 2: Apple's "Don't Build, Just License" AI Strategy May Be Fatally Flawed

The widely praised Dediu/Asymco thesis holds that Apple brilliantly avoided overbuilding AI infrastructure by licensing Gemini, since models will commoditize. Thompson inverts this: in the agentic era, the model-harness integration is the product, and Apple's approach of building its own harness (new Siri) on top of a licensed model may not be competitive.

"Microsoft decided that they couldn't deliver a compelling product by going that route [model-agnostic harness]; what has Apple done to inspire faith that they can do a better job?"

Supporting evidence: Dediu's own cited data shows DeepSeek built a comparable model for $6M vs. $100M, and open-source models power 80% of VC-backed startups — but Thompson argues this misses the point that the integration layer, not the model itself, is the defensible asset.

---

Contrarian 3: Anthropic and OpenAI Are More Durable — Not Less — Than They Appeared Late Last Year

The prevailing concern has been that frontier model companies face existential margin pressure from commoditization. Thompson argues agent harness integration actually strengthens their position, making them the integration point of the value chain rather than interchangeable model vendors.

"If agents are making Anthropic and OpenAI the point of integration in the value chain, then the bubble argument that these companies are overvalued, or that the massive investments other companies are making on their behalf in data centers is unwarranted, may not be correct."

---

3. Companies Identified

---

4. People Identified

---

5. Operating Insights

Insight 1: Enterprises Should Cut Headcount More Aggressively Than Feels Comfortable, Then Rebuild with Agents

Thompson's analysis implies that companies that timidly trim around the edges will be structurally disadvantaged versus AI-native competitors with smaller cost bases and greater agent-driven output. The operational implication is a bias toward aggressive restructuring now, not cautious optimization.

"The most forward-looking and future-proof approach will likely be to cut more rather than less, with the hope that those who remain have no choice but to rebuild scale with agents."

---

Insight 2: For AI Product Builders, the Harness Is the Product — Not the Model

Teams building AI-powered products should invest heavily in the orchestration and agent harness layer, not just in selecting the best underlying model. The differentiation (and pricing power) lives in the integration between model and harness. Shipping a great harness on a good model beats shipping a mediocre harness on a great model.

"What made Opus 4.5 compelling was not the model release itself, but changes to the Claude Code harness that made it suddenly dramatically more useful. What this means is that model performance isn't the only thing that matters: the integration between model and harness is where true agent differentiation is found."

---

Insight 3: Enterprise AI Pricing Should Be Outcome-Based, Not Seat-Based

Microsoft's move to a $99/seat E7 bundle (double the prior E5 tier) illustrates that the enterprise pricing model is shifting to justify productivity outcomes. Operators building enterprise AI products should frame pricing around measurable productivity gain, not per-user access — because that is the value enterprises are actually buying.

"Microsoft is going to bundle AI into a new higher-tiered enterprise offering, E7, which is going to cost twice as much — $99 per seat per month — as the formerly top-of-the-line E5. That's a big increase, which Microsoft needs to justify with AI that actually makes those seats more productive."

---

6. Overlooked Insights

Overlooked Insight 1: CPU Demand (Not Just GPU) Is a Beneficiary of the Agentic Era

Thompson makes a brief but notable point that agent workloads involve more than GPU-intensive model inference — the agent orchestration layer and its deterministic tool use is better handled by CPUs. This is a meaningful and underappreciated signal for the broader semiconductor landscape, pointing toward Intel and other CPU-focused players as agentic infrastructure beneficiaries.

"The agent itself needs compute, and that compute — and the tools the agent uses — is better done by CPUs than GPUs."

---

Overlooked Insight 2: Reasoning Models Have Already Foreclosed the Local Inference Opportunity

Thompson notes in passing that reasoning models — not just agents — have already killed the local/on-device AI thesis, because they require too much memory and compute to run locally. This is a significant strategic dead end for any company (including Apple) betting on on-device AI as a durable advantage in the frontier capability tier.

"Not only do reasoning models require fast compute, given the number of tokens generated, but they also need exponentially more memory to accommodate much larger context windows, which is the biggest limitation of local models... there is also no scenario where capable reasoning models that are remotely competitive with cloud-based models are running locally in the foreseeable future."