Why AI Capex Forecasts Keep Being Too Low

Analysis

Every major AI infrastructure number reported this quarter has landed on the high side of estimates, not the low side -- guidance raised, backlogs bigger than expected, credit lines expanded rather than trimmed. That's unusual for a spending cycle this large and this late into its second year, and it's worth asking what it actually means for anyone pricing a private round off the public comps.

Start with the clearest data point, Broadcom's latest quarter:

Management didn't just beat -- it raised the forward number again rather than easing off it.

Marvell told investors substantially the same story three weeks after its own fiscal second-quarter earnings had triggered a 6% selloff on execution worries. Pulse covered that guidance raise as it happened: a fiscal 2028 custom-silicon revenue target lifted to roughly $18 billion, up from a prior target near $16.5 billion, with the stock rallying more than 6% on the update.

Oracle's backlog tells a similar story at a different scale. The company's remaining performance obligations reached $638 billion at the end of its last fiscal year, up 363% year over year, according to ERP Today's breakdown of the print -- and roughly $75 billion of that is prepaid or customer-supplied hardware, which reduces how much capital Oracle itself needs to raise to build the data centers behind it. Oracle reports its next quarterly results after the close today, and the figure worth tracking isn't the backlog total again -- it's whether the conversion rate from that backlog into recognized revenue is holding up.

The Chipmakers Aren't Alone

The pattern extends past chipmakers:

• Qualcomm -- a new multi-generational chip pact with Amazon carries a potential $60 billion ceiling, with a warrant tying Amazon's upside directly to Qualcomm hitting delivery milestones.
• Anthropic -- close to finalizing a $15 billion pre-IPO revolving credit facility, six times last year's $2.5 billion line, following the same bank-syndicate playbook SpaceX ran before its own record debut.

Banks don't commit billion-dollar tranches to companies they expect to slow down.

None of this is happening in a vacuum. Nvidia remains the dominant supplier of AI training silicon by a wide margin, and Broadcom and Marvell together control an estimated 95% of the custom-ASIC design market underneath it -- so the same underlying hyperscaler capex is what's driving all five of these numbers higher at once, not five independent demand signals. When one number moves, they tend to move together, which is exactly what would make a slowdown, if one ever comes, likely to show up everywhere at once too.

What The Guidance Doesn't Show

What none of these figures answer is how much of this spending converts to profit, or how exposed it is to a small circle of companies buying from and selling to each other. Guidance is not booked revenue, however -- Marvell's own stock fell 6% on a beat-and-raise quarter in August before rallying on a guidance raise three weeks later, proof that even the companies posting these numbers can't fully convince the market they'll hold. Backlog conversion has lagged guidance before across the software industry, and a credit facility is leverage a company has to service whether or not the anticipated revenue shows up on schedule -- the real risk in this whole cycle is a demand air pocket hitting after the debt and capacity commitments are already locked in, not before.

For VCs marking portfolio companies against these public comps, the discipline is separating the guidance number from the cash number in every deal memo that cites one of them. Oracle's earnings print later today is the next test of whether that gap is closing or widening.