Meta has a plan to deploy four successive generations of its own in-house AI chips by the end of 2027, Bloomberg reported this week. That timeline is aggressive. Most chip programs take three to four years just to reach volume production; Meta is proposing to stack four generations inside twenty months. The announcement makes clear that something important has shifted in how the largest AI consumers think about silicon: not as something you buy, but as something you build, tune, and iterate on the way you iterate on software.
The chips in question are variants of Meta's MTIA architecture, which the company has been developing quietly since 2021. The first generation handled recommendation and ranking workloads. The newer variants are aimed squarely at inference: the computational work of running a trained model against billions of user queries every day. That distinction matters. Training is where NVIDIA still dominates and will for some time; the engineering complexity of distributed training across thousands of GPUs is genuinely hard, and NVIDIA's software stack has a ten-year head start. Inference is different. The workload is more predictable, the latency requirements more specific to each use case, and the economics of scale strongly favor purpose-built silicon over general-purpose GPUs rented by the hour.
Meta is not alone. Microsoft's Maia 200 is built for exactly the same inference economics. Google has been running its TPU line through multiple generations for years. Amazon has Trainium. Apple has its Neural Engine. What Bloomberg's reporting adds to this well-known picture is the pace of Meta's ambition: four chips in two years suggests the company is treating its hardware roadmap like a software release schedule, shipping fast and iterating hard rather than waiting for the "perfect" design.
Bloomberg Intelligence put the size of this market in perspective: AI accelerator chips are projected to grow at 16% annually to reach $604 billion by 2033, up from $116 billion in 2024. That is a fivefold increase in under a decade. The report also projects that application-specific integrated circuits, the custom chips that companies like Meta and Google design for their own use, will grow faster than general-purpose GPUs, at a 21% compound annual rate. At those volumes, even small improvements in cost-per-token translate to billions of dollars in savings annually.
The strategic logic has a second dimension beyond cost. Custom chips mean control. When your inference runs on your own silicon, you can optimize the whole stack: the hardware, the software runtime, the quantization strategy, and the model architecture itself. You can test model variants that are specifically shaped for your chip's memory bandwidth rather than being constrained by what runs well on hardware designed for everyone. This matters more as AI becomes core infrastructure rather than an experimental feature. Companies that build on NVIDIA indefinitely are trusting that NVIDIA's roadmap will always serve their needs; companies building their own silicon are betting they know their needs better than any chip vendor can.
There is a risk in all of this that does not get discussed enough. Custom chip programs are capital-intensive, technically demanding, and slow to show returns. The development costs run into the billions. If a generation of chips underperforms, the teams that built it may have wasted a year of training runs and engineering time. NVIDIA's advantage is partly its silicon and partly the fact that it has been making these mistakes and learning from them since the 1990s. The hyperscalers are young in the chip business, and four generations in two years leaves very little room for error. Meta is betting that its engineering culture and its scale are enough to compress a learning curve that took NVIDIA decades to climb. It is not an unreasonable bet. It is also not a guaranteed one.
What is clear is that the market is bifurcating. There will be companies that buy compute from hyperscalers running NVIDIA hardware. And there will be a small number of companies large enough to justify building their own. The gap between those two groups, in cost, capability, and strategic autonomy, is about to get wider.