Microsoft plans to increase production of its own AI chips, Bloomberg reports, a move that signals the company wants a larger share of its accelerator supply under its own control rather than dependent on external vendors.

The report gives no capacity figures, dollar commitments, or manufacturing partners, so the scale of the ramp remains unquantified. What it does confirm is direction: Microsoft intends to push more of its custom silicon into service as demand for AI compute keeps straining the industry's supply of high-end accelerators.

Microsoft has already designed its own AI accelerator family, the Maia line, which the company has discussed publicly as part of its infrastructure strategy. Custom silicon of this type gives a hyperscaler several commercial levers. It can tune the chip to its own workloads, avoid margin paid to merchant vendors, and — critically over the past two years — secure compute capacity that the open market could not deliver at any price.

The supply chain context here is straightforward. NVIDIA's data-center GPUs have been the default accelerator for AI training and inference, and the gap between what cloud providers want and what the merchant market can ship has pushed every major hyperscaler toward in-house designs. Google runs its TPU program at scale. Amazon has its Trainium and Inferentia families. Microsoft's Maia effort is the same play: build your own floor of compute when buying someone else's is the bottleneck.

Manufacturing such chips still requires advanced foundry capacity, advanced packaging, and high-bandwidth memory — the same constrained inputs that make merchant GPUs scarce. Microsoft has not disclosed which foundry or packaging partners would carry the reported production increase, and Bloomberg's report does not name them. That leaves open the question of whether the ramp reflects new wafer allocations secured in advance or a shift of existing allocation toward Microsoft's own designs.

For NVIDIA, the commercial significance is incremental rather than existential. Microsoft remains one of the largest buyers of merchant GPUs for its Azure cloud, and its custom silicon to date has complemented rather than replaced those purchases. But every accelerator Microsoft builds in-house is a unit it does not buy, and hyperscaler self-supply is the structural pressure point analysts watch most closely in the AI chip market.

For Microsoft itself, the calculation is margin and scheduling. Internal chips are cheaper per unit of compute if volumes are high enough to amortize design and software costs, and they give the company scheduling certainty it cannot get from a merchant vendor allocating product across many customers.

The report also lands as the wider AI infrastructure buildout continues to absorb record capital spending across the sector. Cloud providers have repeatedly raised their data-center budgets to keep pace with model training demand, and silicon supply — not just power and land — has been a binding constraint on how fast that spending converts into usable compute.

How quickly Microsoft can convert a production plan into deployed capacity depends on factors the report does not address: foundry allocation, packaging throughput, and the maturity of the software stack that makes custom silicon usable for customers rather than only for internal workloads. Those constraints, not design intent, have set the pace for every hyperscaler's custom-silicon program so far, and they will determine whether Microsoft's planned boost meaningfully shifts its compute cost base or simply adds a few points of self-supply on top of continued merchant purchases.

Bloomberg's report offers no timeline, so investors and competitors alike will be watching Microsoft's next infrastructure disclosures for a concrete figure on how much of its AI compute will run on its own silicon.