Beijing is taking a closer look at Broadcom’s hardware inside Chinese state-backed data centers, according to a Financial Times report relayed by Reuters on September 24, 2026. The timing is awkward for Broadcom in one sense and telling in another: the company just raised its own fiscal 2027 AI-chip revenue forecast to roughly $115 billion, up from a prior call of “more than $100 billion,” and it expects that number to nearly double to $230 billion in fiscal 2028. China’s scrutiny and Broadcom’s growth math are, in effect, two sides of the same story. Custom AI silicon has become central enough to global compute infrastructure that governments now treat it as a strategic asset worth watching, not just a line item on an earnings call.
The past four weeks of chip news read like a single narrative arc. Broadcom’s forecast bump. Qualcomm’s $60 billion custom-silicon pact with Amazon. A $145 million funding round for an obscure interconnect startup called Eliyan that just made it a unicorn. Each of these events, taken alone, is a business story. Taken together, they describe a market where the fight for AI compute has moved well past Nvidia’s GPU business and into a sprawling ecosystem of application-specific chips, warrant-laden supply deals, and geopolitical risk assessment.
What Happened: Beijing Examines Broadcom Hardware in State Data Centers
The Financial Times report, cited by Reuters on September 24, 2026, said Chinese authorities are examining the use of Broadcom hardware deployed inside state-backed data centers. Public reporting available so far does not name the specific agency conducting the review, the data-center operators involved, or which Broadcom hardware line is under the microscope. What is confirmed is the fact of the review itself and its timing: it lands three weeks after Broadcom told investors its custom AI-chip business was accelerating faster than previously modeled.
That combination matters because Broadcom’s AI business is not built on off-the-shelf GPUs sold to whoever can pay. It is built on custom application-specific integrated circuits, or ASICs, designed jointly with a small number of hyperscale customers. A government review of that hardware inside domestic infrastructure raises different questions than a review of, say, imported gaming graphics cards. It touches on data sovereignty, supply-chain dependency, and the broader US-China tension over who controls the compute that trains and runs frontier AI models.
Broadcom’s $115 Billion Forecast, Explained
Reuters reported on September 2, 2026 that Broadcom raised its fiscal 2027 AI-semiconductor revenue forecast to approximately $115 billion, a figure also detailed by Yahoo Finance. CEO Hock Tan had previously guided investors toward AI-chip revenue of about $58 billion for fiscal 2026, a year that ends in October. The new fiscal 2027 figure represents roughly a doubling from that base, and Tan told investors the company expects AI-chip revenue to reach around $230 billion in fiscal 2028, another near-doubling.
That kind of forecast only holds up if customers keep signing multi-year infrastructure commitments, and Broadcom laid out exactly who those customers are. Tan said the company has visibility into more than 10 gigawatts of future AI infrastructure tied to Anthropic, more than 5 gigawatts tied to OpenAI, and 3 gigawatts tied to Meta. Reuters separately named Meta Platforms, Alphabet’s Google, and OpenAI as companies already running Broadcom’s custom AI accelerators and networking silicon in production.
| Fiscal Year | Broadcom AI-Chip Revenue | Change vs. Prior Year | Source |
|---|---|---|---|
| FY2026 (guided) | ~$58 billion | Baseline | CEO Hock Tan, Reuters |
| FY2027 (raised forecast) | ~$115 billion | ~98% increase | Reuters, Sept. 2, 2026 |
| FY2028 (guided) | ~$230 billion | ~100% increase | Reuters, Sept. 2, 2026 |
A separate gigawatt breakdown helps translate those dollar figures into physical infrastructure. A gigawatt of AI data-center capacity is a rough proxy for tens of thousands of accelerator racks, the power substations to feed them, and the cooling systems to keep them running. When Broadcom talks about “visibility” into 10-plus gigawatts for a single customer, it is describing a multi-year buildout, not a single purchase order.
Why Custom Silicon, Not Just GPUs, Is the New Battleground
For most of the current AI boom, the default assumption was that Nvidia GPUs sat at the center of every serious training or inference cluster. That assumption still mostly holds for training frontier models, but the economics of inference, the day-to-day work of actually running a deployed model for millions of users, have pushed hyperscalers toward custom chips designed for their own specific workloads. Broadcom does not sell a branded chip the way Nvidia does. It designs ASICs in partnership with customers, reportedly including Google, Meta, and OpenAI, then leaves the branding and deployment to the customer.
This is a fundamentally different business model than Nvidia’s, and it explains why Broadcom’s stock has become an AI bellwether in its own right, alongside Nvidia’s. When a hyperscaler builds a custom chip through Broadcom, it typically does so because a workload has stabilized enough, say, ranking and recommendation inference, or a specific class of transformer inference, that the flexibility of a general-purpose GPU is no longer worth its cost premium. Our earlier coverage of Google’s Ironwood TPU pricing move against Nvidia’s B200 and B300 traced the same logic: Google Cloud lists Ironwood at $12.00 per chip-hour on demand in Iowa, dropping to $5.40 per hour on a three-year commitment, a structure aimed squarely at customers running steady, predictable inference loads rather than bursty experimentation.
Qualcomm’s Parallel Bet: A $60 Billion Amazon Deal
Broadcom is not the only company converting hyperscaler AI spending into custom-silicon revenue. Reuters reported on September 8, 2026 that Amazon could purchase up to $60 billion of Qualcomm AI data-center chips and related products under a long-term partnership. As part of the arrangement, Qualcomm granted Amazon warrants valued at roughly $4 billion, giving Amazon the right to purchase up to 25 million Qualcomm shares at $161.26 each, with those warrants set to expire on September 3, 2036, according to a regulatory filing detailed in CNBC’s coverage of the deal.
The warrant structure is worth pausing on. Rather than a simple purchase agreement, Qualcomm tied Amazon’s equity upside to actual chip purchases, vesting in tranches as Amazon buys Qualcomm server silicon and related technology. That structure gives Amazon a financial incentive to keep buying, and it gives Qualcomm a way to lock in a marquee cloud customer without discounting the chips themselves. It echoes the compute-for-equity arrangements Nvidia and OpenAI have used over the past two years, now applied to a company, Qualcomm, that has spent most of its history in smartphones rather than data centers.
Inside Qualcomm’s AI200 and AI250
Qualcomm’s data-center push, detailed on the company’s own newsroom page, centers on two chip families, the AI200 and AI250, both announced in October 2025 as the company’s formal entry into AI-inference hardware. The AI200 supports 768 GB of LPDDR memory per card, an unusually large capacity that favors running big models with lower power draw than comparable HBM-based accelerators. Qualcomm targeted commercial availability of the AI200 during 2026, while the AI250, an inference-focused ASIC built around near-memory compute, is aimed at a 2027 launch.
The Amazon agreement does not specify, in public reporting so far, whether it locks in AI200 units, AI250 units, or a later Qualcomm generation. What is clear is the strategic intent: Qualcomm wants to be known as more than a smartphone chip supplier, and inference, running trained models cheaply at scale, is the workload where a memory-heavy, power-efficient design can undercut Nvidia on cost per query even without matching Nvidia’s raw training throughput.
The Interconnect Bottleneck: How Eliyan Became a Unicorn
While Broadcom and Qualcomm chase headline chip deals, a smaller story underscores where the next bottleneck sits. Eliyan, a Santa Clara-based startup building chip-to-chip and chiplet interconnect technology, announced a $145 million Series C on July 29, 2026, valuing the company at $1 billion, according to Eliyan’s own announcement and Reuters. The round was led by Seligman Ventures, with new strategic backing from Cisco Investments and Lumentum, alongside participation from an early investor in Mellanox, the networking company Nvidia acquired years ago.
Umesh Padval, managing partner at Seligman Ventures and a former Mellanox board member, joined Eliyan’s board as part of the round. Eliyan’s technology sits within the UCIe ecosystem, the Universal Chiplet Interconnect Express standard for linking chiplets, memory, and networking components on the same package. The company describes its work as solving the data-movement bottleneck between AI chips, a problem that grows more acute as accelerator clusters scale into the tens of thousands of units that Broadcom and Qualcomm’s customers are now deploying.
The through-line here is simple: as custom silicon multiplies across Broadcom ASICs, Qualcomm’s AI200 and AI250, and Google’s TPUs, the interconnect fabric linking all of it together becomes as economically important as the chips themselves. Eliyan’s jump to unicorn status in a single Series C round signals that investors now see chiplet interconnect as a distinct, investable layer of the AI hardware stack rather than a footnote inside a larger chip design.
Deal Size Comparison: Broadcom, Qualcomm, and Eliyan
| Company | Deal / Forecast | Reported Value | Date |
|---|---|---|---|
| Broadcom | Raised FY2027 AI-chip revenue forecast | ~$115 billion | Sept. 2, 2026 |
| Broadcom | FY2028 AI-chip revenue guidance | ~$230 billion | Sept. 2, 2026 |
| Qualcomm-Amazon | Custom AI chip purchase commitment | Up to $60 billion | Sept. 8, 2026 |
| Qualcomm-Amazon | Share warrants granted to Amazon | ~$4 billion (25M shares at $161.26) | Sept. 8, 2026 |
| Eliyan | Series C funding round | $145 million ($1B valuation) | July 29, 2026 |
Historical Context: From GPU Scarcity to Custom Silicon Sprawl
Two years ago, the dominant AI hardware story was Nvidia GPU scarcity: waitlists for H100s, resellers marking up allocations, and cloud providers rationing access to compute. That scarcity narrative pushed every major hyperscaler toward the same conclusion at roughly the same time, that depending entirely on one supplier for the industry’s most important input was a strategic risk worth paying to avoid. Google had already been building TPUs in-house for years through Broadcom’s fabrication partnership. Amazon had Trainium and Graviton. Microsoft had Maia. What changed over the past 12 months is the scale and public visibility of those bets.
Broadcom’s own numbers tell that story. AI-chip revenue guided at $58 billion for fiscal 2026 was already a substantial business by any historical semiconductor standard. Guiding to $230 billion just two years later assumes the custom-silicon share of total AI compute spending keeps growing relative to merchant GPUs, not just growing in absolute terms. That is a bet on a structural shift, not a cyclical bump, and it is the same bet Qualcomm made by entering data-center chips for the first time with the AI200 and AI250.
Market Impact: What This Means for Nvidia and the Broader Chip Trade
None of this displaces Nvidia’s role in training the largest frontier models. Broadcom’s own ASIC customers still buy Nvidia GPUs for cutting-edge training runs. What it does is cap Nvidia’s share of the inference market, the part of the AI compute stack that scales with users rather than with model size. Every dollar Meta, Google, or OpenAI spends on a Broadcom-fabricated inference chip is a dollar that does not flow through Nvidia’s high hardware margins.
That dynamic also explains why China’s scrutiny of Broadcom hardware carries more weight than a routine customs review would. If Chinese state-backed data centers are running Broadcom-designed silicon, any restriction, export-control tightening, or forced substitution would ripple through both Broadcom’s revenue and China’s own AI infrastructure buildout at the same time. It is the mirror image of the export-control pressure the US has placed on Nvidia chips sold into China for the past several years, this time aimed at the custom-ASIC side of the business rather than merchant GPUs.
Competitive Landscape: How the Custom Silicon Players Stack Up
Broadcom, Qualcomm, and Nvidia are now pursuing overlapping but distinct strategies inside the same fast-growing AI hardware market. Broadcom’s model is design-as-a-service: it partners with hyperscalers to build chips those customers own and brand themselves. Qualcomm is entering data centers as a challenger with a memory-heavy, power-efficient architecture aimed at inference cost. Nvidia remains the default for training, defended by its CUDA software moat, while increasingly facing inference-side competition from exactly the custom chips Broadcom and Qualcomm are building.
Google’s Ironwood TPU sits as a useful real-world data point in this competition. Independent analysis from SemiAnalysis found Ironwood inference cost of about $0.181 per million tokens at a 100 tokens-per-second target, versus roughly $0.222 for Nvidia’s B200 and $0.276 for the newer B300, putting Ironwood 19% cheaper than B200 and 34% cheaper than B300 on that specific benchmark. Numbers like that explain why Broadcom, which helps fabricate custom chips for hyperscalers pursuing the same strategy Google pioneered with TPUs, is projecting revenue growth that outpaces the broader semiconductor industry.
The China Angle: Why This Isn’t Just a Business Story
The September 24 report on Chinese authorities examining Broadcom hardware arrives against a backdrop of years-long US export restrictions on advanced chips headed into China, restrictions that have already reshaped how Chinese firms source AI hardware. If Beijing’s review results in restrictions on Broadcom hardware inside state infrastructure, it would represent a rare case of China restricting an American chip designer’s presence domestically rather than reacting to a US-imposed export limit. Public reporting has not yet detailed the scope, legal basis, or likely outcome of the review, and any additional specifics should be treated as unconfirmed until named Chinese regulators or Broadcom itself comment directly.
What Happens Next: Five Predictions
First, expect Broadcom to face direct questions about the China review on its next earnings call, likely with a carefully worded response that neither confirms nor denies specific restrictions. Second, Qualcomm’s AI200 launch during 2026 will be watched closely as the first real-world test of whether its memory-heavy architecture can compete with Nvidia and Broadcom-fabricated ASICs on cost per inference query, not just on paper specifications. Third, expect at least one more chiplet-interconnect startup to raise a large round in the next two quarters, following Eliyan’s unicorn valuation, as investors chase the layer beneath the chips themselves.
Fourth, Broadcom’s named gigawatt commitments to Anthropic, OpenAI, and Meta will likely expand to include additional hyperscalers or foundation-model labs as those companies lock in multi-year capacity ahead of expected compute shortages. Fifth, and most speculatively, further government scrutiny of custom AI silicon, not just merchant GPUs, is likely to spread beyond China, as more governments treat the physical infrastructure running AI models as a matter of national strategic interest rather than pure commercial procurement.
Frequently Asked Questions
What exactly are Chinese authorities examining regarding Broadcom?
According to a Financial Times report relayed by Reuters on September 24, 2026, Chinese authorities are examining the use of Broadcom hardware deployed inside state-backed data centers. Public reporting has not specified which hardware line, which agencies, or what outcome is expected.
How much AI-chip revenue does Broadcom expect to make?
Broadcom guided to roughly $58 billion in AI-semiconductor revenue for fiscal 2026, raised its fiscal 2027 forecast to approximately $115 billion, and expects that figure to reach about $230 billion in fiscal 2028, according to Reuters reporting on comments from CEO Hock Tan.
Who are Broadcom’s main custom AI chip customers?
Reuters identified Meta Platforms, Alphabet’s Google, and OpenAI as companies using Broadcom’s custom AI accelerators and networking chips, with additional multi-gigawatt infrastructure visibility tied to Anthropic, OpenAI, and Meta specifically.
What is the Qualcomm-Amazon AI chip deal worth?
Amazon could purchase up to $60 billion of Qualcomm AI data-center chips and related products under a long-term partnership announced September 8, 2026. Qualcomm also granted Amazon warrants worth roughly $4 billion, covering up to 25 million shares at $161.26 each, expiring September 3, 2036.
What are Qualcomm’s AI200 and AI250 chips?
Announced in October 2025, the AI200 is a data-center AI chip supporting 768 GB of LPDDR memory per card, targeted for commercial availability during 2026. The AI250 is a follow-on inference-focused ASIC using near-memory compute, targeted for a 2027 launch.
What does Eliyan make, and why did it become a unicorn?
Eliyan builds chip-to-chip and chiplet interconnect technology within the UCIe ecosystem, aimed at reducing data-movement bottlenecks between AI chips in data centers. It raised a $145 million Series C on July 29, 2026, led by Seligman Ventures with Cisco Investments and Lumentum participating, valuing the company at $1 billion.
Does this affect Nvidia’s position in the AI chip market?
Nvidia remains the primary supplier for training frontier AI models, but growth in custom ASICs from Broadcom’s hyperscale partners and new entrants like Qualcomm is concentrated in inference workloads, where cost per query rather than raw training throughput determines the winner.
Is this the same as the US restricting Nvidia chip exports to China?
No. The Broadcom review reported on September 24, 2026 describes Chinese authorities examining American-designed hardware already inside Chinese state infrastructure, a different dynamic than US export controls restricting outbound chip sales to Chinese buyers.
