Last updated on July 23rd, 2026 at 07:52 am

One of the hardest things to acknowledge and a change is adjusting away from something that’s worked for a long time, even when it’s clear that method is no longer working. That is what is happening at Microsoft and Google. Microsoft has typically launched demonstrably flawed products that emulate an existing business model into its vast distribution system, and it over time iterates into effectiveness. Google has basically printed money on search for decades at revenue, funding its business which it has largely spent on a series of ineffective boondoggles (remember Google Glass?) what’s become apparent is that AI is where both of these business models have failed to adapt. Google just raised equity money for the first time in two decades, is launching new models late and is producing millions of inaccurate search results an hour. Microsoft has now changed its AI strategy four times in less than a year, including watching seven of its own models last month, then yesterday announcing a partnership with France-based Mistral. Neither company seems to be making much progress. Let’s look at the problems and then the potential reasons.

Microsoft’s approach started with a partnership born out of conflict that looked like destiny. Billions into OpenAI, a partnership that made Copilot the front end of the most important model company on earth, and for a while the arrangement looked less like a bet than a coronation. Then came the weekend in late 2023 when OpenAI’s board fired Sam Altman, Microsoft appeared to have him and half the company for a few hours, and the whole thing evolved into a far more complex resolution. While the stake and the dependency stayed, and the partners said all the right things, something didn’t work out. The signs were early in Copilot adoption failure. And over the following two years the partnership that was supposed to be the answerbecame the thing Microsoft spent its energy trying to escape.

Copilot’s early stumbles with OpenAI were followed up by the addition of Anthropic’s Claude, then surging, which was then eventually all but abandoned because of token costs. Then a move that made it clear exactly what was driving the strategy: Microsoft began routing the coding work in Copilot off the expensive frontier models it was renting and onto a small in-house model of its own, MAI-Code-1-Flash, priced to escape the per-token economics of paying Anthropic and OpenAI across hundreds of millions of calls. That is a company reading its own income statement and flinching.

Then, in June, the in-house solution. Seven in-house MAI models announced at once, a whole family, headlined by a reasoning model the company benchmarked against the best on the market. It landed on a crowded market like a stone dropped in a well. A month later Mira Murati’s Thinking Machines, a lab of roughly a hundred and fifty people, shipped a single open-weight model called Inkling and collected more coverage in a morning than Microsoft’s seven got in a week. The company with a mid-sized law firm’s headcount made more noise. The three-trillion-dollar company’s announcement all but disappeared.

And now, this week, a multibillion-dollar deal with Mistral. European compute, European sovereignty cover, the French lab’s models threaded into Foundry and Copilot Studio. Structured carefully to avoid the equity stake that drew Brussels’ antitrust attention the last time. A fourth door, opened while the first three still stand ajar. This has all happened in a matter of months.

Watched as a sequence, it looks like panic. A frog on a hot plate, jumping because standing still burns. The more useful reading is that every jump is the same instinct, the instinct that built the company but is no longer serving it.

Grinding, and the time it takes

Microsoft has almost never been first, and has almost always won anyway.

Excel arrived after Lotus 1-2-3. Windows arrived after the Mac had already shown people what a graphical interface was. Teams arrived years into Slack’s head start. Azure arrived after Amazon had effectively invented the cloud business. In every case the pattern was identical and it was devastating. Let someone else prove the category exists and absorb the cost of educating the market. Then bring the distribution, the enterprise contracts, the bundling, the integration into software the customer already runs, and grind. Not a sprint past the incumbent but a long, patient erosion, three years or five or ten, until the day the leader’s head start stopped being worth anything and the switching cost pointed entirely one direction.

That machine has been one of the most effective in the history of business. It also has a single, absolute requirement. It needs the target to hold still long enough to be ground down. The whole method assumes a market that will sit and wait while Microsoft optimizes its way into it, because optimization takes time, and time is the input the machine consumes.

Artificial intelligence moves at a pace that does allow for the grind.

A market that refuses to wait

The frontier now resets on a cadence measured in weeks, even days. In a single stretch this summer the leaderboards absorbed a run of major releases from labs across three countries, and the question of which model is best had a different answer at the end of it than at the start. There is no stationary incumbent to erode, because the incumbent is replaced before the erosion finishes, and the optimization shipped to close last quarter’s gap answers a question the market has already moved past.

This is the argument the first piece in this series made in the abstract. The volume and velocity of change have outrun the ability of even the best-resourced institution on earth to respond by refinement. The playbook needs a pause the market will no longer grant. Microsoft is what the consequences look like, wearing a suit.

Look at where the MAI models actually went and the shape of the problem sharpens. The small, cheap, unglamorous ones shipped and are doing real work. The coding model is in Copilot. The voice, transcription, and image models are in production with public pricing. Those are commodity layers, and a company that owns its own commodity layer saves a fortune at Microsoft’s scale, and none of it requires winning anything. The flagship, the reasoning model, the one carrying the entire claim that Microsoft can build frontier intelligence itself, has spent the weeks since its announcement in private preview, available to select partners, powering nothing you can touch, its numbers self-reported and never once submitted to a public head-to-head. On every leaderboard the industry actually watches, Microsoft does not appear. Not ranked low. Absent.

A company confident in a frontier model enters it in the arena. Microsoft kept the reasoning model behind glass and shipped the autocomplete. The seven-model family, read against what shipped, was one frontier bet the company was not ready to place, packaged at a number large enough to imply a program it was not ready to run.

The reason for the panic pace

When Fortune covered Inkling, it framed the model as filling a void left because the American market is lagging Chinese developers in competitive open releases. The specimen, the neolab, and the pace-setter, in one sentence, in a mainstream financial outlet.

China set this tempo, and it set it by declining to play the game Microsoft is built to win. The strength of the Chinese effort is that it does not orient itself around the American labs at all. It builds its own systems, ships them open-weight, iterates fast, and moves past rather than through. The historical rhyme is the United States after independence, which did not spend the nineteenth century fixated on Britain but built its own industry and railroads and institutions and walked past the empire that birthed it without asking permission. That is the posture now, and it is precisely the posture that a grinding machine cannot answer, because you cannot erode a competitor who is not standing in front of you waiting to be eroded. They are already somewhere else.

Google, and the cost of getting it right

Eighteen months ago Alphabet carried roughly $12 billion in long-term debt and was essentially unleveraged. It is now cash flow negative for the first time in a decade despite its core businesses expanding. It now carries over $102 billion in total debt across six currencies, including a 100-year Sterling note maturing in 2126. On June 3 it priced an $84.75 billion equity capital raise, upsized from a planned $80 billion after demand poured in, anchored by a $10 billion private placement from Berkshire Hathaway, structured as Class A and Class C common stock, two series of mandatory convertible preferred at 6.25%, and a $40 billion at-the-market programme beginning in the third quarter. The stated purpose in the filing is expanding AI infrastructure and compute. Capital expenditure guidance for 2026 runs to $190 billion, with 2027 guided to increase significantly on top of that. Self-fund, borrow, dilute. Alphabet has now walked through all three inside a year and a half, and it took the equity window ahead of the OpenAI and Anthropic raises, absorbing a finite pool of AI-allocable capital before its competitors could reach it.

The infrastructure case for that spending is real. Cloud is carrying a backlog north of $460 billion and Sundar Pichai has said that compute constraints capped revenue that demand would otherwise have delivered. The money is buying something.

It is not buying a flagship model. Google’s last stable frontier reasoning release is Gemini 3.1 Pro, from February. Gemini 3.5 Pro was announced at I/O on May 19 with a June target. June passed. July passed. A widely reported July 17 date passed. Bloomberg reported on July 16 that the release had slipped again after the model fell short of Google’s own internal quality goals, on hallucination rates and real-world reliability, with DeepMind said to have scrapped and rebuilt the base model from the ground up. Prediction markets have moved to late July or early August. Google has reportedly registered a Flash stopgap, and some reporting suggests the Pro release could be skipped altogether. Five months without a flagship, in a market where the frontier resets every six weeks, at $190 billion a year.

You can read a strategy into Microsoft’s flailing, at least partially. Own the commodity layer to protect the margin. Rent the frontier from whoever holds it this month and let them carry the cost and risk of leading. Lease the political positioning from Mistral that you cannot manufacture. That is a coherent supply-chain strategy for a company that has decided the frontier is someone else’s problem to win on Azure’s rented hardware, and that its own job is distribution and margin, the two things it has never lost. The legacy business is enormous and durable. The channel is real.

Google is more complicated, but the reality is the same factor is affecting both. It’s speed. Large companies do not move quickly, enormous ones with decades long, unassailable business models even less so. They are built on protecting, on defence and rarely on offense, but we live now in a world where the attacks are coming at a completely different pace. they are not even on Google or Microsoft core business models, which are essentially intact and this is part of the AI leadership psychosis that we have talked about elsewhere. Google has cannibalized its market, cutting off traffic to millions of sites in order to issue. Millions of wrong answers an hour through AI search. The need to “be there” is so great it is causing some of the most successful companies in history to abandon their very foundations of their success, prematurely, seemingly for no reason.