Owning the Stack Was Always Zoho's Bet. AI Just Made It Obvious.

Owning the Stack Was Always Zoho's Bet. AI Just Made It Obvious.

AI Infrastructure  /  Enterprise Software
Anthropic is exploring building its own chips. That story is about Anthropic. The larger story is about what happens to every software company that does not control its inference costs.
$30B
Anthropic revenue run rate, Apr 2026
3.5GW
TPU compute Anthropic committed to from 2027
~$500M
Estimated cost to design a custom AI chip
16
Zoho-owned data centers globally

Reuters reported this week that Anthropic is exploring building its own AI chips. The plans have no committed design, no dedicated team, and may not happen at all. That is the correct level of caution for a company that has never built silicon. It is also beside the point. The fact that a company generating $30 billion in annualized revenue is even having this conversation tells you where the pressure is coming from.

Chips are not an engineering aspiration. For AI companies running inference at scale, they are a cost survival question.

The math changed this year

Anthropic's revenue run rate went from roughly $9 billion at the end of 2025 to over $30 billion now. That growth compounds the chip problem rather than solving it. More customers means more inference. More inference means more compute. And right now, that compute comes almost entirely from NVIDIA, Google Tensor Processing Units, and Amazon Trainium. None of those are cheap, and none of them are under Anthropic's control.

Earlier this month, Broadcom filed an 8-K confirming a long-term agreement to supply custom TPUs for Google, with Anthropic committed to consuming 3.5 gigawatts of that capacity starting in 2027. For context, Anthropic was consuming around one gigawatt earlier this year. The new commitment triples that before the year is out. That is not a procurement decision. That is a dependency.

Vendor-reported figures Designing a custom AI chip costs roughly $500 million, according to industry sources cited by Reuters. That covers engineering talent and manufacturing validation. It does not cover the time, typically several years, before the chip generates any return. Anthropic has the revenue to consider it. Most software companies do not.

Zoho made this bet before it was obvious

At ZohoDay 2026 in February, Zoho laid out something that did not get enough attention in the coverage. The company is investing in its own server infrastructure and database technologies, not as a side project but as a deliberate cost and sovereignty strategy. Zoho owns 16 data centers globally and runs its own large language model, Zia LLM, trained without routing data through external providers. The explicit position: Zoho is not a compute vendor, and it does not intend to pass GPU costs to customers.

That decision was made years before inference costs became a boardroom conversation. It looked like frugality at the time. It looks like strategy now.

Sridhar Vembu built Zoho to avoid depending on anything he could not control. That instinct turns out to be load-bearing in an AI world.

The Zoho model is not replicable for everyone. It required years of patient capital, a privately held structure with no pressure to show short-term margins, and a founder willing to build data centers in rural Tamil Nadu rather than rent capacity from hyperscalers. Most software companies cannot do that. But the underlying logic, that owning the cost layer is worth the upfront investment if you intend to run AI at scale, is now something every enterprise software company has to reckon with.

This is not just an AI company problem

The companies most exposed are not the frontier model labs. They have the revenue and investor backing to either buy chips in volume or eventually design their own. The companies most at risk are the mid-tier software vendors that embedded third-party AI into their products during 2024 and 2025 and are now discovering that inference costs scale faster than the value they can charge for them.

Andy Jassy made the same point from the other side in Amazon's shareholder letter this week. Amazon built Trainium specifically to reduce its own inference costs, and expects that investment to save tens of billions in capital expenditure annually at scale. The companies that did not make that kind of bet are now renting compute from the companies that did. That is a margin structure problem, and it compounds every quarter.

Viability question for CIOs and CTOs

When your software vendor added AI features in 2024, ask them how they are pricing inference costs two years from now. If the answer is a flat fee or bundled into existing tiers, find out what happens to that pricing when their GPU bills triple. If the answer is consumption-based, you are already absorbing their chip dependency whether you know it or not.

Zoho's answer to that question has been consistent for years: they absorb the cost internally because they control the infrastructure. That is a vendor stability argument worth understanding before you sign a multi-year contract with anyone whose AI strategy depends entirely on rented compute.

Sources Cherney, Max A., and Deepa Seetharaman. "Anthropic Weighs Building Its Own AI Chips." Reuters, 9 Apr. 2026.
Broadcom Inc. Form 8-K. U.S. Securities and Exchange Commission, 6 Apr. 2026.
Bellamkonda, Shashi, et al. "Zoho Focuses on Its Enterprise-Ready Evolution at ZohoDay 2025." Info-Tech Research Group, 27 Mar. 2025.
Jassy, Andy. "CEO Andy Jassy's 2025 Letter to Shareholders." About Amazon, 9 Apr. 2026.
ZohoDay 2026, Austin, TX. Analyst attendance, Feb. 2026.
Disclaimer: This blog reflects my personal views only. Content does not represent the views of my employer, Info-Tech Research Group. AI tools may have been used for brevity, structure, or research support. Please independently verify any information before relying on it.