(Preview) Microsoft’s Plan for Platform Survival, Meta and the Market’s Permission, A Lack of Situational Awareness
Summary
- Microsoft gained 15% and $450 billion in one day as the market rewarded a strategy built around serving, rather than leading the creation of, frontier models. Ben Thompson’s framing: Microsoft keeps OpenAI’s inference business while training capacity moves elsewhere, then positions itself as the trusted enterprise layer between customers and OpenAI or Anthropic.
- The hyperscaler risk gradient runs from Amazon’s physical infrastructure to Microsoft’s exposed software franchise: “The more physical you are, the safer you are. The more digital you are, the more threatened you are.” Google combines frontier ambitions with threatened search; Meta lacks a cloud business but faces a less acute threat from AI consuming users’ time.
- Microsoft’s model is IBM in the 1990s: accept that monopoly-era scale weakened product excellence, then turn breadth, consulting and incumbent relationships into the advantage. As Lou Gerstner’s logic is paraphrased, “What we’re good at is being big” — a playbook that gave IBM another 20 to 25 years.
- Microsoft wants to reduce frontier models to processors: stateless components beneath a Microsoft-provided or customer-built orchestration layer, analogous to Intel or AMD chips beneath Windows. The pitch is defensive but coherent — keep proprietary workflows and metadata away from model companies that otherwise learn “how people actually do work” and may eventually replace the software above them.
- The strategic risk is timing: early technologies reward integration because they are “not good enough,” while modular suppliers tend to win later. Enterprises going directly to OpenAI or Anthropic might accept lock-in yet move so much faster that Microsoft’s multi-model harness becomes a lowest-common-denominator constraint.
- Ben’s own experience sharpened that risk: scaffolding that helped model 5.5 on extra high became counterproductive with 5.6 because the newer model had internalized more planning and review. Meanwhile, the industry’s open letter signals fear, not strength: “You don’t sign open letters because you’re winning,” and NVIDIA could face margin compression if OpenAI and Anthropic become the only meaningful chip buyers.
Deep dive
1. Microsoft’s rally masks the AI threat to its software franchise
Andrew Sharp opens with the market verdict: Microsoft rose 15% in one day, adding $450 billion — Bloomberg’s reported biggest single-day value increase in market history — despite having stepped away from the frontier-model race.
Ben maps the hyperscalers by cloud position, frontier ambition and AI exposure. Amazon sits at the safer extreme because “AI is not delivering my packages”; its e-commerce core is arguably better than ever, its cloud business is doing well, and its comfort with tighter margins may help if compute margins compress. Google has cloud and frontier models but threatened search; Meta lacks cloud, pursues the frontier and is less directly exposed than Microsoft.
The durable rule is physical versus digital exposure: “The more physical you are, the safer you are. The more digital you are, the more threatened you are.” Microsoft has Azure, but its productivity software can increasingly be recreated; Ben and Andrew’s team already replaced a narrow Teams-based podcast workflow with software they made themselves.
2. Microsoft is replaying IBM’s survival strategy
Ben’s analogy reaches back to IBM in the 1990s, when investors wanted the conglomerate broken apart. Lou Gerstner’s contrarian answer was that the pieces would struggle independently: “We’re not actually good at anything.… What we’re good at is being big.”
IBM converted breadth into 20 to 25 extra years by becoming the company enterprises trusted to navigate the internet, building consulting capacity and middleware for websites and e-commerce. Microsoft can similarly combine its sales force, forward-deployed engineering and decades-long customer relationships.
The uncomfortable premise is that a former monopoly may have “lost the ability to be great” after becoming “fat and flabby.” Microsoft therefore wins less through singular product leadership than by being acceptable across everything — precisely why Nadella’s earlier shift from Windows toward cloud and services mattered.
Nadella’s unusually personal control of the Build keynote suggested urgency to Ben: “We’re actually in bigger trouble than people realize.” Technology threats take years to reach reported results — as Intel’s problems took roughly a decade to manifest — but by the time the damage appears in the results, “it’s way too late.”
3. The real battle is ownership of enterprise workflow data
Ben remains skeptical that models truly generalize: “I don’t see evidence of generalization yet.” But that question might be moot because enough task data, feedback and verifiable outcomes can teach systems to perform increasingly broad categories of work.
Using Claude Code, Cowork, Codex or another model-company harness exposes more than prompts. Even if training data is excluded, the provider sees which tools workers use, what succeeds and what fails — “the crown jewels of your company,” because that metadata describes how the business actually operates.
Microsoft’s counterproposal is to make models resemble Intel or AMD processors beneath Windows: send each model only the context required for one computation, receive the result, and keep orchestration in a Microsoft-provided or customer-built harness. Microsoft wants stateless models; OpenAI and Anthropic are “trying to be everything.”
4. Integration could beat Microsoft before modularity arrives
Ben calls Microsoft’s strategy logical, especially for enterprises that want model choice and do not want to surrender proprietary context. But Clayton Christensen’s framework cuts against it initially: when technology is still inadequate, integrated solutions win because controlling more layers improves performance.
Microsoft benefits from the argument that models are very expensive and workloads require routing among cheap and costly options. Ben says that narrative felt “a little astroturfed,” because model capability can be more of a ceiling than a floor: a highly capable model can still answer an easy question quickly.
Andrew notes that enterprises may ultimately choose between paying Microsoft and paying OpenAI or Anthropic as a platform. Ben’s counter-scenario: a company going all-in with a frontier provider could accept lock-in yet accelerate so much faster that a cautious rival’s products simply “suck compared to yours.”
The unresolved issue is whether the harness itself matters. Microsoft benefits if models absorb most intelligence and orchestration becomes interchangeable; it loses if its controls withhold useful context, over-engineer the workflow and leave customers “not even coming close” to the models’ capabilities.
5. Model progress is erasing scaffolding while incumbents signal fear
Ben’s concrete example began with 5.5 on extra high, supplemented by a “Superpowers” skill package enforcing planning and review. With 5.6, which already overdid those behaviors, the same scaffolding became “out of control”: the improved model had internalized work previously performed by the harness.
A harness called Pie offered the corresponding design principle: do the minimum and add only necessary tools. Otherwise, work built around today’s limitations becomes harmful three months later because “your harness is competing with your model.”
The industry’s open letter reads to Ben as evidence of alarm over OpenAI and Anthropic’s uptake. Even OpenAI signed — especially revealing because it does not want open-model competitors — but the broader signal is categorical: “You don’t sign open letters because you’re winning.… You sign open letters ’cause you’re freaking terrified.”
NVIDIA’s participation reveals the economic stakes. If only two companies — OpenAI and Anthropic — meaningfully buy chips, NVIDIA becomes the supplier being ordered around, hyperscalers become financing vehicles, and margins compress; if successful startups can be replaced within a year or two, Ben asks whether a durable startup ecosystem remains at all.