Build vs Buy · AI CHIPS · October 9, 2026
Good morning, Folks! Several of the largest technology companies design their own AI chips. Should anyone else? This episode continues our Build vs Buy series.
Custom silicon is the deepest form of "build" in the AI stack, and the one most often attempted for the wrong reasons.
The decision hinges on scale, workload stability, and the software ecosystem the chip has to live in.
What "Custom AI Chip" Means
Design from scratch — You architect an accelerator, work with a foundry, and own the roadmap. Extremely expensive and slow.
Semi-custom with a design partner — A chip partner helps with design and manufacturing for your specific workload.
Use a cloud provider's custom silicon — You rent chips a hyperscaler designed, without designing anything yourself.
Buy merchant GPUs — Purchase general-purpose accelerators from established vendors.
The CODEW Lens: For nearly every company, the real alternatives are the last two.
The Build Case
Designing chips can make sense for organizations running very large, stable workloads where even small efficiency gains compound into large savings; where a specific workload can be served better by a specialized architecture than by general-purpose GPUs; where reducing dependence on a constrained supply chain has strategic value; and where the company already has deep hardware and software engineering depth.
When Building Wins — Three Conditions:
1. Extreme, stable scale. Enough volume to amortize design cost across years.
2. A well-understood workload. Specialization pays only if the workload will not shift under you.
3. Full-stack capability. Hardware, compilers, and software teams working together.
The CODEW Lens: Custom silicon is a bet that your workload will stay put longer than the chip takes to build.
The Buy Case
Buying wins for almost everyone else. Merchant accelerators arrive with mature software stacks, broad framework support, and a large pool of engineers who already know how to use them. Design projects take years, carry manufacturing and supply-chain risk, and can end with a chip that is outpaced by the market by the time it ships. The software ecosystem is often the deciding factor: a faster chip with immature tooling can cost more in engineering time than it saves in hardware.
The CODEW Lens: The chip is the visible half. The software ecosystem is the half that decides adoption.
The Risks of Building Silicon
Time risk — Design-to-deployment cycles are long, while AI workloads change quickly.
Cost risk — Design, validation, and manufacturing commitments are large and mostly front-loaded.
Software risk — Compilers, kernels, and framework support must be built or ported.
Supply risk — Advanced manufacturing capacity is limited and contested.
Talent risk — Chip architects and compiler engineers are scarce.
The CODEW Lens: A late chip is not a cheaper chip. It is an expensive one that arrived after the market moved.
Four Paths Compared
| Factor | Design Own | Semi-Custom | Cloud Custom Silicon | Merchant GPUs |
|---|---|---|---|---|
| Time to deploy | Slowest | Slow | Fast | Fastest |
| Upfront cost | Highest | High | Low | Low to moderate |
| Software maturity | You build it | Partly yours | Provider-supported | Most mature |
| Control | Highest | High | Low | Low |
| Best for | Hyperscale operators | Very large specialized users | Cost-sensitive cloud users | Most enterprises |
The CODEW Lens: Most companies capture the benefit of custom silicon by renting it, not designing it.
A Practical Middle Path
Benchmark your own workloads — Test them on merchant GPUs and cloud custom chips before assuming specialization pays.
Keep software portable — Frameworks that target multiple hardware types reduce switching cost.
Revisit at scale — If spend becomes very large and stable, the build case may open up later.
The CODEW Lens: Earn the right to build silicon through scale first.
Build vs Buy: The CODEW Verdict
Six Questions — In Order:
1. Is your AI compute spend large and stable enough to amortize a multi-year chip program?
2. Is your workload well understood and unlikely to change shape?
3. Do you have hardware, compiler, and software engineering depth in-house?
4. Have you benchmarked cloud custom silicon and merchant GPUs against your workloads?
5. Can you absorb a delayed or underperforming chip?
6. Is supply-chain independence worth the cost and complexity?
The CODEW Lens: Rent specialization, buy generality, and design silicon only when scale makes the chip your strategy.
The Build vs Buy Glossary
Accelerator — A chip built to speed up AI computation.
ASIC — An application-specific chip designed for a narrow set of tasks.
Compiler — Software that translates models into instructions a chip can run.
Foundry — A manufacturer that fabricates chips designed by others.
Merchant GPU — A general-purpose accelerator sold on the open market.
Software Ecosystem — Libraries, tools, and developer skills surrounding a chip.
FAQ
Q: Why do large tech companies design their own chips?
Their scale lets small efficiency gains compound, and they have the engineering depth to support the software.
Q: Should a typical enterprise design its own AI chip?
Rarely. Renting cloud custom silicon or buying merchant GPUs delivers most of the benefit.
Q: What is the biggest risk?
Time. A chip that ships late may already be outpaced by what the market offers.
Q: Does the software ecosystem really matter that much?
Yes. Immature tooling can erase hardware savings through engineering effort.
The CODEW Stat
4 paths · 6 questions · 1 ruleDesign, semi-custom, cloud custom, or merchant. Six questions sort the decision. One rule closes it: build silicon only when scale makes the chip your strategy.
Reviewed by Erwin Castro
on
Friday, October 09, 2026
Rating:
No comments: