We are a Microsoft Partner that publicly chose Claude, in an analysis Forbes cited while Microsoft’s stock was sliding 36%. So you might expect us to squirm now that Microsoft has posted the largest single-day market value gain in stock market history, adding close to half a trillion dollars after its Q4 results. We are not squirming. We like what we see.
Because the numbers behind the surge tell a story most of the coverage missed, and it runs through Microsoft Foundry, a platform most business owners have never heard of, and open weight AI models, a category most business owners are already using without knowing it. Here is what actually happened, and what it means for how you buy, govern, and run AI in your business.
The headline numbers from Microsoft’s Q4 FY26 results were strong on their own. Quarterly revenue of $90 billion. Azure growth accelerating to 43%, with guidance of 45% next quarter. Azure passing $100 billion in annual revenue for the first time. Microsoft 365 Copilot passing 30 million paid seats.
But the number that tells you where this is going is smaller. Foundry, Microsoft’s platform for building AI applications and agents, reached 100,000 customers and more than doubled its revenue year on year. Customers running at a rate of one trillion tokens per year quadrupled. And the number of customers building with models from multiple providers is up five times since the start of the year.
Read that last one again. Microsoft’s own earnings call celebrated customers not standardising on one AI model. Satya Nadella pitched model choice as the strategy: over 11,000 models in the catalogue, including OpenAI, Anthropic, Mistral, xAI and Microsoft’s own MAI family, with the agent and platform layer staying stable while the model underneath swaps based on cost, quality and compliance.
The market did not reward Microsoft for owning the best model. It rewarded Microsoft for making the model swappable. That is the whole story of the record day, and it confirms the position we took in February when we chose Claude as our AI platform: the model is a component, the platform and governance around it are the asset.
One more detail from the results that we found genuinely funny. Microsoft disclosed a $3.2 billion gain on its investment in Anthropic, the maker of Claude. The company that sells Copilot made billions last quarter from betting on the competition. If Microsoft hedges its model bets, your business probably should not be a purist either.
Microsoft Foundry is the shelf, not the product. It is Azure’s platform for building AI applications: a catalogue of models, the tools to connect them to your data, and the infrastructure to run them, all behind one endpoint and one bill.
At Build 2026, Microsoft added Foundry Managed Compute, which lets businesses deploy open source models from the Hugging Face ecosystem onto managed GPU capacity without running their own servers, Kubernetes clusters or model runtimes. Models in the curated catalogue are security screened, ship in a safe weight format with no untrusted code paths, and you can fine-tune them or bring weights you trained elsewhere.
In plain terms: the operational excuse for avoiding open models is gone. Frontier models from OpenAI and Anthropic, and free open weight models from anywhere, now sit on the same shelf, with the same identity controls and the same invoice. Which brings us to the part of this story that matters most for governance.
An open weight model is one whose trained parameters, the “weights”, can be downloaded and run by anyone, on any hardware, usually at no licence cost. You pay for the computing, not the model. We recently priced the practical options in our guide to whether your business should run its own AI model, so we will not repeat the cost maths here.
What has changed since is scale. Vercel, which routes tens of trillions of AI tokens a month between production applications and model providers, reported that open weight models processed 29% of all tokens on its gateway in June, up from roughly one ninth in April, while representing under 4% of total spending. They cost about one tenth of the average token price. Businesses are quietly routing routine work to cheap open models and reserving premium models for work where accuracy matters.
That is rational. It is also where the risk enters, because the open weight leaderboard is not being led by American labs.
In the last week of June, Chinese models accounted for 48% of traffic on OpenRouter, a widely used model routing platform, up from 20% a year earlier, according to CNBC. US models fell from 74% to 32% over the same period. DeepSeek alone handled 22.6% of token volume on Vercel’s gateway in June, making it one of the largest model providers in production use, full stop.
Here is the part that should concern Australian business owners. This adoption is not happening because your staff downloaded DeepSeek. It is happening because the tools they use embed these models under the hood. The clearest example: Cursor, one of the world’s most popular AI coding tools, built its Composer 2 model on Kimi, developed by China’s Moonshot AI. The US Congress is now jointly investigating enterprise exposure to Chinese models, with letters already sent to Cursor and Airbnb.
And the regulatory risk runs both directions. Analysts are warning of a Huawei rerun, where businesses that standardise on a Chinese model get abruptly cut off by Western restrictions. Meanwhile China’s Ministry of Commerce is reportedly considering export controls on model weights, which would strand anyone who built on downloads. A model that is free today and unavailable tomorrow is not free. It is a liability with a delayed invoice.
This is shadow AI’s second act. In February we reported that our audits consistently find 40% to 60% of knowledge workers using unsanctioned AI tools. The first act was staff pasting data into consumer chatbots. The second act is harder to see: approved-looking tools with unapproved models inside them. Your AI register now needs to track not just what tools your business uses, but what models those tools run on, and in which jurisdiction. That is precisely the vetting work our AI governance service exists to do.
The same Vercel data shows where the premium end of the market sits. Anthropic captured 61% of total spending on the gateway while processing only 32% of tokens, with its strongest position in coding, back office automation and application generation, the work where getting it wrong costs real money.
That matches how we deploy AI for clients through our Managed AI service. Claude remains our primary platform for exactly the reasons we set out in February: zero data retention on commercial plans, enterprise identity integration, and depth on document and compliance work. Open weight models have a legitimate place, on-premises workloads, cost-sensitive volume tasks, and sovereignty requirements where the weights need to live inside your boundary. But they are a deliberate architectural choice with governance attached, not a default you inherit because a vendor swapped models to protect their margin.
You will probably never log into Microsoft Foundry. But it changes three things about how you buy AI.
Model choice is now the norm, not the exception. The biggest software company on earth just told investors its customers use multiple models per workload. Any AI strategy that begins with “we are a [single vendor] shop” is already dated. The right question is which model for which task, under what controls.
The switching costs moved up a layer. If models are swappable, lock-in lives in the platform, the data plumbing and the agents built on top. That is where your leverage in vendor negotiations sits, and where your due diligence should focus.
Cheap models will keep arriving inside your existing tools. Vendors are under margin pressure and open weight tokens cost a tenth of frontier tokens. Expect more of your SaaS stack to quietly switch models underneath you. Contracts should require disclosure of model providers and hosting jurisdictions, and your AI programme should verify it.
Ask your AI tool vendors which models they run and where. Send the question in writing to your five most-used AI-enabled tools. If a vendor cannot or will not answer, that silence is your answer. Congress is asking American companies this exact question right now; you are allowed to ask it too.
Add model provenance to your AI acceptable use policy. Most policies written in 2025 govern tools. Update yours to govern models: approved providers, approved jurisdictions, and a review trigger when a vendor changes its model supplier.
Get a shadow AI discovery done before you make platform decisions. You cannot choose a model strategy without knowing what is already running in your environment. Our AI governance onboarding maps every AI tool in use, including the ones with Chinese open weight models embedded. Contact us on 1300 EPIC IT to book one.
Microsoft Foundry is Azure’s platform for building AI applications and agents. It offers a catalogue of more than 11,000 models from providers including OpenAI, Anthropic, Mistral and xAI, plus open source models via Hugging Face, all behind a single endpoint and bill. It reached 100,000 customers and more than doubled revenue in Microsoft’s FY26 results.
Open weight models are AI models whose trained parameters can be downloaded and run by anyone, usually free of licence fees. You pay for computing rather than the model itself. They cost roughly a tenth of frontier model pricing per token, which is why they now process around a third of enterprise AI traffic on major routing platforms.
They carry risks most businesses have not assessed: regulatory exposure in both directions, uncertain behaviour inherited from base model training, and the possibility of abrupt access loss, similar to what happened with Huawei. The bigger issue is that many businesses use them unknowingly, because popular tools embed Chinese open weight models under the hood. A shadow AI discovery identifies where they sit in your environment.
Microsoft’s Q4 FY26 results beat expectations, with Azure growth accelerating to 43%, revenue of $90 billion, and Copilot passing 30 million paid seats. The stock rose 15.5% on 30 July, the largest single-day market value gain in stock market history. Growth in Microsoft Foundry and its multi-model strategy were central to the result.
Sometimes. They suit cost-sensitive volume tasks, on-premises workloads, and sovereignty requirements where model weights must stay inside your boundary. They are a poor fit for high-stakes work where accuracy matters most, which is why premium models still capture the majority of enterprise AI spending. Treat them as a deliberate architectural choice with governance attached, not a default.
