The most important battle in artificial intelligence may eventually have less to do with who builds the smartest model and more to do with who owns the silicon underneath it.
Mumbai: Meta, Microsoft, Google and Amazon are accelerating development of custom AI chips, attempting to reduce costs, secure computing capacity and gain greater control over infrastructure that has become almost as strategically important as the models themselves.
This does not mean Nvidia is suddenly in trouble. Quite the opposite. Nvidia’s latest numbers make the word “challenger” look almost optimistic: the company generated $96.2 billion in quarterly revenue in Q2 FY2027, including $89 billion from its Data Center business, up 117% year-on-year.
But Big Tech has clearly decided that renting the engine forever is less attractive than eventually knowing how to build one.
Meta Wants Silicon Designed Around Meta
Meta’s answer is its Meta Training and Inference Accelerator, or MTIA, a family of chips developed specifically for its own workloads.
The company says it already operates hundreds of thousands of MTIA processors for recommendation and advertising inference. Its MTIA 300 is in production, while MTIA 400, 450 and 500 are being developed for broader generative-AI inference workloads through 2027. Meta is targeting four new chip generations within two years.
The financial incentive is not subtle.
Meta currently expects 2026 capital expenditure of $130 billion to $145 billion, largely as it expands infrastructure for its AI ambitions. At that scale, shaving even a modest percentage from compute costs stops being an engineering curiosity and starts looking rather attractive to the finance department.
Custom silicon also gives Meta something money cannot always immediately buy: optimisation. A processor designed specifically for recommendation, ranking or inference does not need to behave like a general-purpose accelerator built for everyone.
There is, naturally, a catch. Designing chips is difficult. Manufacturing them at scale is worse. Meta still relies on partners across semiconductor design, fabrication, memory and networking. “In-house” silicon, like many modern miracles, comes with quite a few outside contractors.
Microsoft Is Building Its Own Economics
Microsoft has taken a similar path with Maia.
Its latest Maia 200 AI accelerator is designed primarily for inference, with Microsoft claiming more than 30% better performance per dollar than existing systems in its fleet. It is being deployed for Microsoft’s AI workloads including Copilot, Foundry and internal model development.
Microsoft expects approximately $190 billion in capital expenditure during calendar 2026, including around $25 billion attributed to higher component pricing. The company has acknowledged that capacity constraints are expected to persist through 2026.
That makes Maia about more than competing with Nvidia. It is an insurance policy against scarcity.
The strongest AI companies increasingly need multiple sources of compute: Nvidia GPUs where they make sense, AMD hardware where appropriate, and internally designed silicon when economics or availability demand it.
Amazon And Google Are Already Further Down The Road
Google has arguably demonstrated the custom-chip thesis better than anyone. Its Tensor Processing Units now span multiple generations, with its latest architecture including TPU 8t for training and TPU 8i for agentic AI workloads. Google says TPU 8t systems can scale to clusters exceeding one million processors.
Amazon has its own increasingly serious silicon business through Trainium and Graviton. Trainium and Graviton together have surpassed a $10 billion annual revenue run rate, while Amazon says 1.4 million Trainium2 chips have already been deployed. Project Rainier alone uses more than 500,000 Trainium2 processors.
Amazon expects roughly $200 billion of capital expenditure in 2026, much of it connected to AWS and AI infrastructure.
And the diversification continues. On September 8, Amazon and Qualcomm announced plans to jointly develop multiple generations of custom data-centre chips, with potential business agreements reaching $60 billion over a decade.
Apparently owning one chip strategy is no longer considered sufficiently ambitious.
Nvidia’s Real Defence Is Bigger Than A GPU
The negative case for Nvidia is clear: every workload shifted to MTIA, Maia, Trainium or TPU represents compute that might otherwise have required Nvidia hardware.
The positive case is considerably stronger than that sentence suggests.
Nvidia’s advantage includes CUDA software, networking, systems engineering, developer adoption and years of optimisation, not merely the chip itself. Custom accelerators also tend to target selected workloads rather than replacing general-purpose GPUs everywhere.
So this is unlikely to become Nvidia versus everybody else.
It is becoming Nvidia plus everybody else.
For Big Tech, custom AI chips provide greater control, potentially lower operating costs and protection from supply bottlenecks. The downside is billions in investment, manufacturing complexity, and the uncomfortable possibility of designing hardware for AI workloads that evolve before the silicon even arrives.
The AI chip race is therefore not really about dethroning Nvidia tomorrow.
It is about making sure that five years from now, Nvidia is not the only door into the building.
Read More: Cybersecurity’s Next Bold Upgrade
