Amazon CEO Andy Jassy says chip business is tripling every year

Amazon CEO Andy Jassy says the company’s chip business is growing threefold every year. For a company best known for selling everything from paper towels to cloud servers, that is a striking claim.

The figures behind it are just as notable. Amazon’s custom silicon division has gone from a side project to one of the fastest-growing parts of the company, and Jassy thinks most people are still underestimating it.

The numbers behind the tripling

Jassy has laid out the growth in 2026 shareholder communications and on earnings calls. Early in the year, the silicon business crossed an annual revenue run rate of more than $20 billion.

That figure was later updated to over $25 billion. Amazon described the expansion as triple-digit growth year over year.

A quick translation for anyone who doesn’t speak finance: a run rate takes current revenue and projects it across a full year. It is a snapshot of pace, not a tally of money already banked.

Jassy has also argued that the headline figure undersells the business. His case is that if Amazon’s silicon unit operated as an independent company, its revenue run rate could reach approximately $50 billion a year.

The gap between $25 billion and roughly $50 billion comes down to accounting perspective. Much of the value of Amazon’s chips shows up inside AWS, where customers rent computing power rather than buy chips outright.

That growth has positioned Amazon as one of the top three data-center chip businesses in the world.

Who is buying, and how much

The demand side of the ledger is where things get large. Amazon has secured more than $225 billion in multi-year commitments for Trainium, its line of AI chips.

Those commitments come from notable customers, including Anthropic and OpenAI.

Adoption is also broad on the more general-purpose side. Amazon says 98% of AWS’s top 1,000 customers on EC2, its rentable server service, are now using Graviton5 technology.

On the AI front, the Trainium2 and Trainium3 generations are nearing full subscription. Amazon attributes that demand to price-performance gains of 30-40% compared with previous GPU offerings.

Sold out, and spending heavily

Capacity is now the constraint. AWS’s AI and compute capacity is sold out through much of 2027 and into 2028.

Amazon is trying to address that. The company has planned capital expenditures of $220 billion for 2026, a figure that covers the data centers, power and hardware needed to keep up.

Why Amazon builds its own chips

For AWS, that means designing processors tuned to its own data centers rather than relying entirely on third-party chipmakers. Graviton handles general computing tasks, while Trainium targets the heavy lifting of AI model training.

The approach turns Amazon into both a customer and a competitor of the chip industry. It still buys hardware from outside vendors, but every workload that shifts to Graviton or Trainium is one that doesn’t need someone else’s silicon.

What this means

For the semiconductor industry, Amazon’s numbers are a signal worth tracking. Companies selling AI and data-center chips now face a hyperscaler that is also a supplier, backed by more than $225 billion in Trainium commitments.

For AI developers, the picture is mixed. Cheaper compute is welcome, but sold-out capacity through much of 2027 and into 2028 means access, not just price, is becoming the bottleneck.

Long-term commitments like those from Anthropic and OpenAI effectively reserve space that smaller players may struggle to secure.

For Amazon itself, the silicon unit adds a new storyline alongside retail and cloud. A business that Jassy suggests could be worth an approximately $50 billion run rate on a standalone basis is no longer a supporting act.

The $220 billion capex plan for 2026 is enormous, and the payoff depends on demand staying as strong as current bookings suggest.

The next markers will come in future earnings calls. Investors and competitors alike will be watching whether the run rate keeps climbing past $25 billion, whether the tripling pace holds, and whether Amazon can build capacity fast enough to match the orders already on its books.