This is the third of five parts. Part one covered what to watch when Kimi K3's weights drop on July 27; part two laid out the financial reality of the AI companies and how K3 acts on it. Part three takes the market's day of judgment two days later — Microsoft's fiscal Q4 2026 earnings on July 29, after the US market close. One other report sits inside this same fortnight: news that broke on July 17 that Meta is in talks to lease compute to Anthropic. K3's weight release (the model going free), Meta's lease (an entry into selling infrastructure), and Microsoft's earnings (the market's verdict) — reading those three along a single line is the view this piece takes.
Where the Market Already Is — The Doubt Is Priced In
Start with the premise: Microsoft has already been sold. The stock is down roughly 20 percent in 2026 and sits nearly 30 percent below its high. The forward P/E has fallen to 19, the lowest since 2016. The center of the market's doubt is capital expenditure: $190 billion for the fiscal year ending June 2026, up 60 percent year over year. Nothing has been shown about when or how that outlay comes back as profit.
So the fork on the 29th is clear. Expectations are low, and if an upside surprise in operating margin or progress on AI monetization appears, there is room for a sharp rebound off the cheapness — one bullish call has the stock up 50 percent if the P/E returns to its customary level near 30. On the other side, if the capex guidance for fiscal 2027 issued alongside the results shows another increase, the decline resumes; that crash risk has been flagged too. That much is the consensus on the numbers. What this piece wants to add is the question to ask beyond them.
What Does the Bull Case Assume?
For the P/E to revert — for the 50-percent scenario to work — one story has to hold. The story is that AI expands Microsoft's margins. The enormous capex is front-loaded investment; once Copilot and Azure's AI revenue mature, the high gross margins of a software company return. That future is what the bulls are buying and what the bears are selling doubt about.
Which is why what to actually watch on the 29th is not revenue growth. It is where gross margin goes, and whether that story continues to hold.
The Economics K3 Introduced
This is where Kimi K3 enters. The 2.8-trillion-parameter open model Moonshot AI released on July 16 sits just behind the top closed models in independent evaluations, with full weights promised for the 27th. And Microsoft has been reported to be evaluating K3 for Copilot workloads.
At first glance this looks good for Microsoft. Microsoft has no frontier model of its own; it is a reseller that sources models from OpenAI and Anthropic. Swap a paid API for free weights and Copilot's cost of goods falls. A rescue for the cost problem — that is how it reads.
But lower costs do not mean higher profits, because the price ceiling is set from outside. If the substance is the same K3, Moonshot itself sells the API at $3 in and $15 out, other clouds can host the same weights, and enterprises hold self-hosting as an option. In the resale of a commoditized input, a fall in cost does not become margin; it gets eaten by price competition. When your wholesale price drops, so does the shop next door's — the economics of a gas station. Evaluating K3 may extend Copilot's life, but it is not a path to higher profit. And what the bull case needed was precisely a story about higher profit.
Meta's Lease — Infrastructure Commoditizes Too
On July 17 it was reported that Meta had entered early talks to lease compute to Anthropic: up to $10 billion over two years, paid monthly, with early termination available to both sides, at Anthropic's approach in June. If it closes, it would be Meta's first large deal selling its own AI infrastructure externally; the company has stood up a division called Meta Compute and brought in a former AWS executive to build the business. Meta's 2026 capex runs to as much as $145 billion.
The report can be read three ways, and all three bear on the 29th.
Read from the supply side, an advertising company is entering the compute rental market. Add a new lessor carrying $145 billion of capex and the price of inference and training infrastructure gets more competitive still. What is happening in this fortnight, in other words, is a double commoditization: K3 turning the model and Meta's entry turning the infrastructure from products into generic goods. Azure's AI gross margin is squeezed from above (free models) and from the side (added supply).
Read from the demand side, it confirms the borrower concentration described in part two. The first outside customer Meta found was not a broad market but a single company, Anthropic. A substantial share of hyperscaler AI revenue continues to be, in the end, the spending of these two firms. And Anthropic is spreading its procurement across AWS, Google, NVIDIA, SpaceX, and now Meta — the shape of a large buyer starting to make sellers compete, evidence that pricing power is migrating to the buy side. When power moves upstream, the middle — cloud and resale — loses margin.
Read from the shape of the contract, it is more unsettling. Two years, monthly, mutually cancellable early: this is now the shape of demand that the largest buyers want. For a data center built with five- to twenty-year debt, a customer who is short-term, substitutable, and free to leave is the worst possible shape of demand. The repayment source for the tower of debt described in part two is being replaced by contracts like this one.
Three Things to Watch, Through K3's Lens
First, the fiscal 2027 capex guidance. Two days after frontier-class weights were given away free, Microsoft will explain an investment plan exceeding $190 billion. The question of what that money is buying acquires specificity because K3 exists. There is one escape route in how they answer: rephrase it as "the model can belong to anyone, we sell inference infrastructure," and the spread of open models becomes a demand tailwind for Azure — 1.4TB of weights only runs where someone owns racks. But that rephrasing is also a declaration that they are surrendering the story of the model as a moat.
Second, Copilot disclosure. How much do they quantify seats and growth rates, or decline to? And any signal on price — if discounting begins, or dissolution into the M365 bundle (effectively making it free), that is an admission that no margin remains in resale. Note that the July 1 M365 price increase was justified by making Copilot features standard. How long bundling can defend unit prices against commoditizing substance is the open question.
Third, the OpenAI-related accounting. OpenAI is expected to lose $14 billion in 2026, and Microsoft has been booking its proportional share of that loss. In a world where an open model of K3's class has closed to within months, the reason to keep swallowing enormous losses as the price of exclusive frontier access grows thinner. The very reporting that Microsoft is evaluating K3 functions as a negotiating card against OpenAI.
The Final Criterion
Now the conclusion that runs through all five parts. The condition under which the bull case vanishes at a stroke can be specified precisely: a Sonnet-class model that runs on 128 to 192GB of unified memory — a desktop box costing under $7,000.
The reason is where that band sits. The volume of enterprise practice — code, documents, business systems, everyday decision support — is processed not at the frontier tier but in this middle band, and the mass of the revenue is here too. The moment this band drops into the box, volume that was metered becomes fixed cost (the depreciation of the box and the electricity), and the price ceiling falls to the floor. Three layers collapse at once. Copilot's resale margin. The model companies' middle-band revenue — which is also the pillar under OpenAI and Anthropic. And the rent flowing to data-center SPVs: if middle-band inference escapes to boxes at the edge, the repayment source turns from rent into the purchase price of a box, and the tower of debt from part two is not saved.
The clock on that condition starts running on the 27th. K3 was trained from the outset on the premise of MXFP4 quantization — frontier weights with deployment cost designed in — and once released it becomes a teacher for distillation. Against ceilings of roughly 200B total parameters at 128GB and 350B at 192GB, one round of teacher-supervised distillation is a matter of months, not years. Even the ferocity with which the US government named K3 as a product of distillation and invoked sanctions can be read, at bottom, as a defense of this threshold.
The July 29 Checklist
Five things. Gross margin, actual and explained — has AI entered the margin line? The number in the fiscal 2027 capex guidance. The granularity of Copilot disclosure and any signal on price. How management addresses — or avoids — open models (K3) and Meta's entry into external sales, the double commoditization of model and infrastructure. And the amount booked for OpenAI-related losses. Whichever way the stock moves, these five will surface as the material for judging whether the story that AI expands margins is alive or dead.
Set the questions from part one — do K3's weights and license ship on schedule, is there any sign of a small-model line — against these five in the same week, and the structure laid out in part two (the revenue is real, but the tower of debt stands on borrower concentration) will show which way it has begun to tilt. The answers can be checked on the morning of the 30th.
A disclosure: this piece was developed in dialogue with an Anthropic model (Claude). Anthropic is a party that would be shot in the revenue by this piece's own criterion (a Sonnet-class model dropping into a box), and a competitor to Microsoft. Read it with that discounted.
Related
This is the third of a five-part series.
- Blog What to Actually Watch When Kimi K3's Weights Drop on July 27 — part one: the checklist for the 27th, and the empty tier
- Blog The Reality of AI Companies, and What Kimi K3 Does to It — part two: the financial reality this piece assumes, and the tower of debt
- Blog Three Reckonings Are Closing In, All in the Same Week — part four: three limits outside AI arriving the same week
- Blog IT Ends as an Industry, and the Age of Materials and Assembly Begins — part five: the same criterion, seen from the side of industry
- Blog Three Transitions in Software — the structural map behind this series
References
- Meta in talks with Anthropic over up to $10 billion in AI compute leasing (BUSINESS+IT) — https://www.sbbit.jp/article/cont1/186164
- Meta in talks to lease up to $10 billion in compute to Anthropic; would be its first external-sales deal (XenoSpectrum) — https://xenospectrum.com/meta-anthropic-compute-lease-talks/
- Microsoft faces crash risk into July 29 earnings; capex outlook in focus (IG Securities, 2026-07-22) — https://www.ig.com/jp/news-and-trade-ideas/Microsoft-202604-06-earnings-report-preview-260722
- Microsoft announces fiscal Q4 2026 earnings date (BigGo Finance, 2026-07-08) — https://finance.biggo.jp/news/ir_MSFT_20260708_21a347e98844
- Prediction: Microsoft Stock Is Going to Soar After July 29 (The Motley Fool, 2026-07-16) — https://www.fool.com/investing/2026/07/16/prediction-microsoft-stock-going-soar-after-july/
- Prediction: Microsoft Stock Could Go Parabolic After July 29 (The Motley Fool, 2026-07-14) — https://www.fool.com/investing/2026/07/14/prediction-microsoft-stock-could-go-parabolic-afte/
- Run Kimi K3 Locally — Weights July 27 Prep (explainx) — references reporting on Microsoft's evaluation of K3 for Copilot — https://explainx.ai/blog/kimi-k3-run-locally-open-weights-desktop-july-2026
- Kimi K3 popularity strains GPUs; new subscriptions suspended (ITmedia, 2026-07-21) — https://www.itmedia.co.jp/news/articles/2607/21/news099.html
- US government names Moonshot AI, alleging Kimi K3 distillation and GB300 use (XenoSpectrum) — https://xenospectrum.com/moonshot-kimi-k3-distillation-accusation/
- Sam Altman's OpenAI is burning billions (European Business Magazine) — OpenAI's 2026 loss outlook — https://europeanbusinessmagazine.com/sam-altmans-openai-is-burning-billions-most-users-pay-nothing-as-anthropic-closes-in/
- Kimi K3 Open Weights Drop July 27: The Developer Prep Guide (byteiota) — MXFP4, 1.4TB, deployment requirements — https://byteiota.com/kimi-k3-open-weights-july-27-developer-prep/