Two Chip CEOs Just Said the Quiet Part Out Loud: Money — Not Chips — Is What Could Stop the AI Boom


The bottleneck moved from the fab to the bank

For two years, the AI industry’s defining anxiety was physical: not enough GPUs, not enough memory, not enough power to run the data centers. Every earnings call, every keynote, every policy hearing circled the same question — can we build the machines fast enough?

On Monday, in Singapore, the two men who sell the machines said the machines aren’t the problem anymore. Speaking at DBS’s inaugural Asia Future Forward Conference — a day-long AI summit that drew roughly 700 senior business leaders, investors, and entrepreneurs — Broadcom president and CEO Hock Tan told the room that access to capital, not chip, memory, or power shortages, could become the AI buildout’s biggest constraint. Intel CEO Lip-Bu Tan, on the same panel, agreed — and put a clock on it.

“What I think might be the biggest constraint is capital,” Hock Tan said, according to CNBC’s reporting from the conference. Lip-Bu Tan’s version: “If there’s any challenge in the front, it’s capital and the cost of capital, and then the other part is in the longer run, people want to see the ROI.”

Read that second sentence again. The CEO of Intel — a company whose entire business depends on the AI buildout continuing at full sprint — is publicly saying investors will eventually want to see returns. That’s not a doom prophecy. It’s a grace period with an expiration date.

Why this matters: the bills are already signed

The Tans weren’t theorizing. The commitments are already on the balance sheets, and they’re enormous. Anthropic’s IPO prospectus — filed earlier this month, and the most candid document the AI industry has ever produced — disclosed $518 billion in non-cancelable compute bills. That’s not a budget forecast. Non-cancelable means the checks are owed whether the revenue shows up or not.

Stack that against the industry’s cash flow. OpenAI corrected its annualized revenue figure to roughly $50 billion last week — a real number, but one built partly on cloud-partner accounting that counts gross rather than net. The frontier labs are spending at a pace their revenue doesn’t yet cover, and the gap is bridged by debt, equity, and long-dated infrastructure deals.

Hock Tan made an important distinction here. He pushed back on the idea that OpenAI is overvalued, telling the panel that the problem is investors setting extreme expectations “without understanding the technology,” per Hong Kong’s The Standard. His actual confidence was in the revenue side: he said Anthropic and OpenAI have shown they can generate significant revenue from their latest models, the ones being deployed as AI personal assistants. The demand is real. The question is whether it’s real fast enough to service the financing.

The shortages the industry obsessed over are easing — on schedule

The second half of Hock Tan’s message was the good news nobody quoted as loudly. The physical bottlenecks are resolving, roughly on the timeline the industry projected:

  • Memory: “Memory shortage today might not be that way a year from now.”
  • Wafers: Chip wafer shortages could ease by 2028 or 2029.
  • Power: Supply should improve as new infrastructure comes online.

But he paired that with a sobering footnote: “There is more demand for compute infrastructure than we can supply,” and that gap could persist through 2029, possibly into 2030. The industry has spent two years overcoming one supply bottleneck after another — and it did. Which is precisely why the constraint moved. When you can finally get the chips, the question stops being can you build and starts being can you afford to keep building.

Hock Tan also named the two scenarios that could break the demand side of his forecast: businesses adopting advanced AI models more slowly than expected, or the models themselves stopping their rapid improvement. Either would leave the industry holding a decade of infrastructure bills against a flattening revenue curve.

The two-to-three-year grace period

Lip-Bu Tan’s contribution was arguably the most important line of the panel. He expects AI demand to stay strong over the next two to three years — the adoption of AI applications and the hunger for compute are both real, and neither is fading. But that “two to three years” is doing a lot of work. It’s the window in which the industry’s bet has to start paying for itself.

The concern, as reported by CTech, points to “a potential shift in the debate surrounding AI. After years in which the industry’s most pressing challenge was securing enough computing capacity to meet demand, the ability to finance that expansion, and eventually generate adequate returns, is becoming a central question.”

This is the same tension that has been building all quarter. The White House’s voluntary AI safety accord is gone — the FTC opened its first enforcement probe of AI agents last week, ending the self-regulation era. Model prices are collapsing at the low end — Anthropic’s Claude Haiku 5.5 matched GPT-6 Luna at $0.10 per million input tokens to the cent. Cheap inference is great for adoption but terrible for anyone whose business model assumed scarcity pricing. And the open-weights flank — Mistral’s trillion-parameter “Le Chonk,” Reflection AI’s coming release — keeps putting pressure on the closed labs’ margins.

Every one of those trends makes the ROI question harder: cheaper models grow the market but compress the revenue per unit of compute, while the compute bills are fixed and non-cancelable.

What this means for builders and buyers

The Tans’ warning isn’t a sell signal — both were explicit that the boom has years of runway. But it changes the shape of the next phase, and there are practical takeaways:

For enterprise buyers: If your vendor’s pricing depends on infrastructure financed at 2026 borrowing costs, your contract is downstream of the capital markets. Diversify vendors, negotiate exit clauses, and watch for pricing resets in 2027–2028 — that’s when today’s financing structures start facing the ROI question. The cheapest model today may be the one whose builder can’t afford to keep it cheap.

For AI startups: The window for “build it and raise your way out” is narrowing. Lip-Bu Tan’s clock is a message to the ecosystem: revenue per unit of compute is becoming the metric that matters, not parameter counts or benchmark scores. Architect for cost-per-token now — the Fireworks/Ember-1 lesson, that efficient reasoning cuts the real bill, applies to product design as much as to model training.

For investors: Hock Tan’s rebuke — that expectations are being set “without understanding the technology” — is worth taking seriously. The gap between “AI is transformative” and “this specific company earns its valuation” is where the capital-constraint story bites. Demand for compute through 2029 is a bullish case for infrastructure. It is not automatically a bullish case for everyone selling picks and shovels at today’s multiples.

For everyone else: Watch the financing announcements more than the model announcements. The next phase of the AI race will be decided less by who has the smartest model and more by who can afford to keep training them. The labs know this — it’s why Anthropic is going public with a $2 trillion ask and why OpenAI’s unit economics are under a microscope. The Tans just said it out loud.

The honest version of the story

There’s something refreshing about this panel. For years the industry’s story was that every constraint was temporary — the chips would come, the power would come, the models would keep improving. Now the two CEOs closest to the physical buildout are saying: the physical part is fine. It’s the money.

That’s actually the more honest framing of where AI stands in late 2026. The technology works — agents are in production, personal assistants are generating real revenue, enterprise adoption is maturing from pilots to measurable returns in places like Singapore. The open question was never really “does AI work.” It’s “can the economics of building it outpace the cost of financing the buildout.” And the people who would know best just told 700 executives in Singapore that the answer is: probably, but the clock is running.