Broadcom's fiscal Q3 2026 earnings, reported September 2, are the cleanest evidence yet that the AI chip story isn't Nvidia versus everyone else. Revenue hit $29.6 billion, up 86% year over year; AI semiconductor revenue alone surged 221% to $16.7 billion, on demand from six customers building custom chips.^[1] Broadcom raised its own fiscal 2026 AI revenue guidance to $58 billion, and laid out $115 billion for fiscal 2027 and $230 billion for fiscal 2028 — guidance that implies AI chip demand keeps roughly doubling for two more years running.^[1] None of that revenue is Nvidia's. It's the design-partner business behind chips Nvidia doesn't make at all.

The chip inside the frontier labs' own designs

Broadcom is the common design partner behind three separate "labs building their own silicon" stories that usually get told as unrelated news. It designed OpenAI's first custom chip, Jalapeño, unveiled in June 2026 and headed for deployment in late 2026 — a purpose-built inference processor that strips out training logic entirely to optimize for running (not building) large language models, manufactured by TSMC with Celestica handling system integration. Early testing shows it beating Nvidia's Blackwell systems on inference-efficiency benchmarks.^[2][3] Broadcom also designs Meta's MTIA accelerator, built specifically for inference and recommendation workloads at scale, and ships Google's TPUs to Anthropic — which is on track to become Broadcom's single largest custom-chip customer in 2027, deploying 1 gigawatt of TPU capacity in 2026 and another 5 gigawatts in 2027.^[1]

The pattern underneath all three deals is the same: training still runs mostly on Nvidia's general-purpose GPUs, because flexibility matters most while a lab is still figuring out what works. Inference — running a finished model billions of times a day — is massive, repetitive, and exactly the workload a narrow, purpose-built chip beats a general-purpose GPU on for cost and power. The pivot away from Nvidia isn't happening in the part of the business Nvidia built its reputation on. It's happening in the part that comes after.

Nvidia's own moat may not cover the customers it's chasing

Nvidia's advantage has always rested on CUDA — the software stack that makes its GPUs the path of least resistance for anyone writing AI code. But that moat isn't uniform across Nvidia's customer base. Frontier labs have the engineering depth to write their own compiler stacks for whatever silicon makes economic sense: Google has XLA, Meta has its own MTIA toolchain, OpenAI now has one built with Broadcom for Jalapeño. They don't need CUDA's convenience. The customers who actually can't afford to leave — smaller AI shops, researchers, enterprises without a dedicated ML infrastructure team — are the ones without the engineering headcount to rewrite a PyTorch/TensorFlow/cuDNN stack for different hardware.

AMD and Intel aren't in the same position, even though both get filed under "Nvidia alternative"

AMD's MI300X already has real, production frontier-tier deployments: Meta runs it for Llama 3 and Llama 4 inference across its infrastructure, Microsoft Azure has it powering production Azure OpenAI Service workloads, and Sam Altman has confirmed OpenAI's own research and GPT models run on MI300 through Azure.^[4][5] That's the concrete case that already fits the "AMD has a real opening" theory. Intel's Gaudi accelerators have had far less traction with any frontier lab so far.

Intel's real structural advantage isn't a chip that wins a given benchmark — it's that Intel owns its own fabs. AMD and Nvidia are both fabless, entirely dependent on TSMC for every advanced-node wafer, competing with each other and with Apple and Qualcomm for the same limited allocation. Intel fabs on its own lines in Arizona, Ohio, and Ireland, is pursuing its own next-generation 18A and 14A nodes, and is opening that capacity to outside customers through Intel Foundry Services. TSMC's advanced-node capacity is the actual physical bottleneck constraining the whole industry — part of why Nvidia's own supply has been constrained — and Intel not needing to wait in that line is durable regardless of which specific chip wins a given round.

The one family relationship in this story runs through the rivalry, not the partnership

Nvidia's Jensen Huang and AMD's Lisa Su are blood relatives — Huang's mother is a sister to Su's grandfather, a relationship genealogists have mapped as first cousin once removed, though both have described it more loosely as "some complex second cousin type of thing." Both were born in Tainan, Taiwan. Neither grew up knowing the other; they met as adults, already well into competing careers.^[13] That's the one family tie anywhere in this piece, and it sits between the two CEOs actually racing head-to-head for the same GPU customers.

The relationship that isn't blood at all is the one doing more real work. Jensen Huang and Supermicro's Charles Liang have known each other since founding their companies the same year, 1993, and Nvidia picked Supermicro as its original go-to-market partner for data center GPUs — a relationship Liang has described as running "almost since day one." Nvidia now accounts for roughly 71% of Supermicro's revenue and 64% of its component spend.^[14] But even that closeness has real leverage running one direction: Supermicro has no long-term supply contract with Nvidia, only purchase orders it renegotiates on Nvidia's terms — and Supermicro's own cofounder, Wally Liaw, was arrested in March 2026 on charges of illegally exporting Nvidia-powered servers to China, a case that cost Supermicro a third of its market value in a week.^[14] A thirty-year friendship built the partnership. It didn't erase the power imbalance underneath it, or insulate either company when a cofounder's own conduct became the story.

Run the logic backward and the inversion is the point, not a footnote. Blood is supposed to be the thing that explains a business alliance — it's the classic mechanism in immigrant-entrepreneur networks specifically, where family is often exactly how a first customer or first partner gets found. Here it explains nothing: Huang and Su didn't know each other existed as relatives until both were already running rival companies, and the relationship never became a business one. The alliance that actually moved billions of dollars a year had no family tie to draw on at all — just two men who happened to start companies in the same year and kept talking for three decades. The tie you'd bet on predicts nothing. The one you wouldn't is the whole partnership.

GlobalFoundries isn't trying to win this race — it's running a different one

The obvious question is why GlobalFoundries, a major US foundry, hasn't captured any of this business. The answer isn't capacity — it's a deliberate strategic exit. GlobalFoundries chose years ago to stop chasing the leading-edge nodes (5nm and below) that frontier AI accelerators require, and it isn't trying to get back in. Instead, its actual AI-adjacent business is running on two different, adjacent fronts. In July 2026 it signed a letter of intent for a $300 million CHIPS Act award to build out silicon photonics — optical interconnects between AI chips rather than the chips themselves, targeting 400 gigabits per second at a fifth of current energy draw, with the federal government taking roughly a 1% equity stake as part of the deal.^[6] And on September 2, the same day Broadcom reported its quarter, GlobalFoundries announced customer availability of its new "UX" chip platforms (40nm and 22nm) built explicitly for "intelligent edge, connectivity, and Physical AI applications" — chips for robots, sensors, and IoT devices, not the frontier models Broadcom, AMD, Intel, and Nvidia are fighting over.^[7] GlobalFoundries isn't losing the AI chip race. It picked a different one.

The fourth race nobody's counting, and it doesn't care who wins the other three

On September 4, SanDisk jumped 11.9% in a single session, closing near $1,740 a share, after Dell's own earnings call flagged memory as its single biggest supply bottleneck — "DRAM, DRAM, DRAM, then NAND, NAND, NAND," in the words of Dell's own operations leadership.^[8] Dell's numbers back up why that matters: fiscal Q2 revenue of $47.0 billion, up 58% year over year, and an AI server backlog that nearly doubled in a single quarter to $95.0 billion, after booking $60.9 billion in new AI server orders in the same three months.^[9] Supermicro, the AI-server builder more exposed to this than anyone — AI GPU platforms are over 90% of its revenue, versus a fraction of Dell's diversified business — is living the same bottleneck from the opposite side of the market, currently holding roughly 40-45% of the AI server segment against Dell's fast-growing share.^[12] The two companies are direct rivals for whose box a GPU ships inside. Neither can out-compete their way around a DRAM shortage.

Micron CEO Sanjay Mehrotra has said he expects the demand to outlast 2027 on "structural supply constraints," and the company has now locked in 16 strategic customer agreements running through 2030.^[10] This is the piece none of the compute-chip competition touches: memory demand scales with how much AI infrastructure gets built overall, not with which vendor's chip goes inside it. Broadcom, Nvidia, AMD, and Intel are fighting over which chip does the computing. Micron and SanDisk don't have a stake in who wins that fight — every outcome still needs the same DRAM and NAND behind it. It's also not a new story for this site: Micron broke ground on its own $15 billion Boise fab in August, built explicitly to help close the gap in US semiconductor manufacturing capacity — that fab is now ramping directly into the demand cycle it was designed for.^[11]