Broadcom expects $56 billion in AI revenue for fiscal 2026 and more than $100 billion in fiscal 2027, figures that position the company as the primary counterweight to Nvidia at hyperscalers. Marvell, the second chip-shepherding player, is tracking toward $11.5 billion in revenue for its fiscal 2027 and $16.5 billion for fiscal 2028.

The scarce commodity behind this shift is not silicon but engineering teams capable of steering complex chip designs through fabs, packaging, and system integration. Hyperscalers and cloud builders are increasingly building their own AI accelerators, and Broadcom and Marvell — the "shepherds" that co-design and deliver those chips — are capturing the resulting volume.

Nvidia co-founder and CEO Jensen Huang said, after expanding a partnership with Marvell, that the company would be the next $1 trillion market cap chip company. Wall Street reacted by selling Broadcom, which had only affirmed its guidance, though Broadcom stock recovered days later. Broadcom already approaches a $2 trillion market cap; Marvell has passed more than a quarter of the way to $1 trillion.

Why do hyperscalers design their own chips?

The economics are straightforward. A homegrown accelerator costs roughly 40 percent of the retail price of an Nvidia or AMD GPU, according to the analysis, and cloud builders can rent it out at 60 to 70 percent of the price they charge for GPUs — keeping 20 margin points in their own pockets to fund further AI spending.

If Broadcom and Marvell keep delivering, somewhere north of half the flops in hyperscaler datacenters will eventually come from homegrown and shepherded chips, extending a pattern already established with homegrown CPUs.

What did Broadcom's Q2 fiscal 2026 deliver?

In the quarter ended in the first week of May, Broadcom posted:

• Revenue of $22.19 billion, up 47.9 percent year on year and 14.9 percent sequentially
• Operating income of $10.79 billion, up 85.1 percent
• Net income of $9.31 billion, up 87.5 percent

The Semiconductor Solutions group drove the growth with just over $15 billion in sales, up 78.5 percent, on Tomahawk, Trident, and Jericho switch ASICs plus AI XPU shipments. Hyperscalers are spending heavily on 102.4 Tb/sec Tomahawk 6 chips, available in plain, co-packaged optics, and scale-up-oriented Tomahawk 6 Ultra variants. Chip operating margins ran at a blended 62 percent; estimated AI chip margins sit in the 50 to 55 percent range, with much of that coming from networking silicon rather than XPUs.

The Infrastructure Software group posted $7.18 billion in sales, up 8.8 percent, with operating income of $5.67 billion — a 79 percent operating margin. VMware likely contributed $5.28 billion in sales, up 15 percent, with operating income above $4 billion, up 32.6 percent, after aggressive cost cuts.

Broadcom ended the quarter with $19.63 billion in cash against $64.91 billion in debt, the residue of the $67 billion VMware acquisition. The company is hoarding cash to secure HBM memory allocations and its place in line at Taiwan Semiconductor Manufacturing Co. — a defensible stance given TPU customers Google and Anthropic and Meta Platforms ramping its MTIA accelerators.

To reach $56 billion in AI revenue, implied AI chip sales must run roughly $16 billion in Q3 and $21 billion in Q4. HBM allocations will likely determine the final number more than any other factor.

How does Marvell compare?

Marvell joins the S&P 500 in two weeks, replacing Campbell Soup Company. Its Q1 fiscal 2027, also ended in early May, showed:

• Revenue of $2.42 billion, up 27.6 percent

• Operating income of $339 million, 14 percent of revenue

• Net income of $34.5 million, down by more than an order of magnitude as costs rose

• Cash of $3.84 billion, up 4.3X year on year, against $4.96 billion in debt

The Datacenter group, now 4.5X larger than the rest of the company, grew 27.2 percent to $1.83 billion. Communications & Other rose 28.7 percent to $585.1 million and remains the profitable counterweight to lower-margin AI XPU work.

Marvell's AI business runs about one-tenth of Broadcom's and leans more on networking than compute. AWS anchors both segments, with Nvidia gaining ground in optics. Estimated Q1 breakdown: roughly $500 million in AI XPU sales — dominated by AWS Trainium chips, up 71 percent year on year — and about $916 million in electro-optical components from the Inphi business, up 47.5 percent. Other datacenter sales likely fell 20.7 percent to $419 million.

The rebalancing from single-vendor compute toward a mix of bought and designed engines looks structurally inevitable — provided both shepherds keep delivering to hyperscalers, cloud builders, and now AI model builders. As the analysis puts it: economic substitution is an economic law, not just a good idea.