Broadcom will guide at least $25 billion of quarterly AI chip revenue on December 9

Bottom line.

1. Broadcom's AI semiconductor revenue rose from $8.4 billion to a guided $21.7 billion within fiscal 2026, and its chief executive said it had secured the supply to roughly double AI revenue in 2027.
2. Custom chips and Nvidia are growing at the same time, and both say supply rather than demand is the limit, so 2027 market shares will reflect who secured wafer, packaging and memory capacity.
3. A first-quarter guide of $25 billion or more on December 9 would put Broadcom slightly ahead of a smooth path to that doubling.

Planning assumption. On December 9, 2026, Broadcom guides first-quarter fiscal 2027 AI semiconductor revenue of at least $25 billion (55 percent: slightly better than even).

Each of the four largest US hyperscalers, Amazon, Microsoft, Alphabet and Meta, now designs its own AI chips as well as buying Nvidia's. Broadcom, which co-develops and supplies many of those chips, including Meta's MTIA and several generations of Google's TPUs, guided $21.7 billion of AI semiconductor revenue for the quarter ending in early November and reports on December 9. This forecast puts a 55 percent probability on its guidance for the following quarter being at least $25 billion, slightly ahead of a smooth ramp to its plan to roughly double AI revenue in fiscal 2027.

Two businesses growing at once

Nvidia remains far larger. Its data center revenue was $89.0 billion in the quarter ended July 26, up 117 percent on a year earlier, and it guided total revenue of $108 billion for the next quarter (Nvidia 10-Q; earnings call transcript). On the same call it gave a preliminary outlook of about 70 percent revenue growth for fiscal 2028, adding: "This is a supply-constrained outlook."

The custom-chip business has a different shape. Broadcom's AI semiconductor revenue, a figure that includes its AI networking chips as well as custom accelerators, was $8.4 billion, $10.8 billion and $16.7 billion in the first three quarters of fiscal 2026, and it guided $21.7 billion for the fourth (Q1, Q2 and Q3 releases). According to a third-party transcript of the September call, chief executive Hock Tan said: "In 2027 we have secured the supply to again double AI revenue to approximately $115 billion" (Benzinga transcript). He also said demand exceeded that outlook and that Anthropic "is on track to become our largest XPU customer in 2027" (ChannelWeb). The $115 billion figure is not in Broadcom's filed release, and Broadcom has not posted an official transcript.

The buyers have each said what they are building. Microsoft said on October 12 that it will scale its Maia chips "to multi-gigawatt deployments starting in 2027", making Maia "the primary inference system for Microsoft products and services", and that Maia 300, built on TSMC's N2P process, "will begin production deployments in 2027" (Microsoft). Meta has scheduled its MTIA 450 for mass deployment in early 2027 and MTIA 500 for mass deployment in 2027, both developed with Broadcom (Meta). OpenAI and Broadcom agreed in 2025 to deploy 10 gigawatts of OpenAI-designed accelerators by the end of 2029 (OpenAI), and OpenAI separately committed to about 2 gigawatts of Amazon's Trainium (OpenAI). Anthropic announced "multiple gigawatts" of TPU capacity with Google and Broadcom coming online from 2027 (Anthropic). Amazon says its own chips business has "exceeded a $25 billion annual revenue run rate" (Amazon).

Google has gone further and started selling its chips outside its own cloud. Alphabet's latest 10-Q says it has "signed a limited number of agreements to supply TPU systems" for customers' own facilities, with "the significant majority" of that revenue to be recognised in 2027, and its inventory rose from $2.4 billion to $10.0 billion in six months, consisting primarily of TPU-system hardware for sale and devices such as Pixel phones (Alphabet 10-Q).

The base rate: a large customer leaving a merchant supplier

The class of situation is a large, technically capable customer replacing a dominant merchant chip supplier with its own design. Such moves usually take longer than planned and, in data centers, have not removed the incumbent entirely. Apple announced in June 2020 that it would move the Mac from Intel to its own chips "in about two years" (Apple) and declared the transition complete in June 2023 (Apple), about a year late. Amazon launched its Graviton server CPU in November 2018 (AWS); by December 2025 it said Graviton had powered more than half of new AWS CPU capacity "for the third year in a row" (AWS). Google has run TPUs in its data centers since about 2015 (Google) and still deploys Nvidia's Vera Rubin racks (Nvidia release). Not every attempt works: Tesla disbanded its Dojo chip team and, according to people familiar with the matter cited by Bloomberg, plans to "increase its reliance on external technology partners" (Transport Topics, Bloomberg).

This case differs from those in one way that matters for December. Apple and Amazon largely controlled their own timelines. Here, both sides say they are limited by the same suppliers. TSMC chief executive C.C. Wei said in October that "the gap between demand and supply are still quite big" (TSMC transcript), and that CoWoS, its advanced packaging, remains customers' first choice. Nvidia raised its supply and capacity commitments from $119 billion to $279 billion in one quarter. Broadcom's growth in 2027 therefore depends less on whether customers want its chips than on how much wafer, packaging and memory capacity it has locked in, and Tan's wording, "secured the supply", is the claim the December guidance tests.

The most likely course, November to December

Nvidia reports its fiscal third quarter on November 17, against the $108 billion guidance. Amazon's re:Invent conference runs from November 30 to December 4 and is where Trainium news usually lands. Broadcom then reports its fiscal fourth quarter on December 9, the date it gave on its September call.

The arithmetic of a doubling sets the bar. Roughly $115 billion over four quarters averages about $28.75 billion a quarter. Starting from $21.7 billion, a smooth path that adds up to $115 billion would put the first quarter at roughly $24 billion to $25 billion; a guide well below that would imply a back-loaded ramp. Broadcom's recent quarterly steps were 29 percent, 54 percent and, on guidance, 30 percent, so a 15 percent step to $25 billion is modest against that record. In the main scenario Broadcom reports fourth-quarter AI revenue at or slightly above its $21.7 billion guidance, reaffirms the fiscal 2027 plan, and guides the first quarter at $25 billion or more.

The consequences reach beyond Broadcom. A guide of that size tells the market that custom accelerators are arriving on schedule, which supports the 2027 capital-spending plans that hyperscalers give in late January and early February; Alphabet chief financial officer Anat Ashkenazi said on its July call that capital expenditure will "increase significantly in 2027" (transcript). It does not imply that Nvidia shrinks. About $40 billion of Nvidia's $89 billion in data center revenue came from what Nvidia calls ACIE (neoclouds, enterprises, industrial and sovereign buyers), customers chief executive Jensen Huang says "don't buy custom chips", though Nvidia said some of that growth came from hyperscalers renting neocloud capacity. On the call Huang described Nvidia as "a platform … that you can use in any cloud", in contrast to many XPUs built "for one cloud or one service." Nvidia is also moving into the parts of the data center that custom accelerators do not replace: it said its Grace CPU revenue topped $5 billion over twelve months, and it expects its CPU revenue to more than double in fiscal 2028.

The rivals, and what would let them win

A first-quarter guide of at least $21.7 billion but below $25 billion is the main rival, at 38 percent. Broadcom has said its 30 gigawatts of customer plans need not all be deployed within the period, and new chip generations often start slowly: Microsoft's previous Maia design had its mass production pushed from 2025 to 2026, according to The Information (Business Today). If the Anthropic and OpenAI volumes are concentrated in the second half of fiscal 2027, Broadcom could reaffirm $115 billion and still guide the first quarter below $25 billion. That outcome would be a timing issue, not a sign that the doubling has failed.

A first-quarter guide below $21.7 billion, or no AI guide at all, is the second rival, at 7 percent. It would most likely mean a supply problem, in packaging, HBM or a new process node.

Dated register

These secondary predictions are published with the forecast and will be checked against the sources named; only the headline claim is scored automatically.

1. Nvidia's fiscal third-quarter 2027 revenue, reported November 17, 2026, is at least $108 billion (85 percent).
2. Broadcom's fiscal fourth-quarter 2026 AI semiconductor revenue is at least $21.7 billion (85 percent).
3. Nvidia's fiscal fourth-quarter 2027 results, around late February 2027, keep or raise the outlook of about 70 percent revenue growth for fiscal 2028 (70 percent).
4. Microsoft states by December 31, 2027 that Maia 300 is in production deployment (70 percent).
5. Meta states by June 30, 2027 that MTIA 450 is in mass deployment (55 percent).
6. Alphabet discloses a dollar figure for TPU-system product revenue in a filing by August 15, 2027 (40 percent).

What to watch

Broadcom's first-quarter AI guide comes first. After it, the split of Nvidia's data center revenue between hyperscalers and other buyers shows where Nvidia's growth is coming from; hyperscalers were $49 billion of the $89 billion last quarter. The third indicator is whether Alphabet starts disclosing TPU-system revenue as its own line, and the fourth is the gap between gigawatts announced and gigawatts reported running at Maia, MTIA and OpenAI's Jalapeño.

The fact that would most change the probability is not public. Broadcom has never published how its 2027 volumes are spread across the year. The doubling target describes the year as a whole, but the December guide covers only its first quarter, so a back-loaded ramp could miss $25 billion while the annual plan stays intact.

What this means

For technology leaders buying AI capacity, the practical consequence is a two-supplier market for the first time: custom accelerators inside the largest clouds, and Nvidia everywhere else. Software that can move between accelerator types keeps that choice open; code written for one cloud's chips gives it up.

For investors, revenue share and chip share are likely to diverge. Custom accelerators can win units and gigawatts while Nvidia's revenue per gigawatt rises, so the clean test is each supplier's reported revenue against its own guidance, not announced gigawatts.

For anyone tracking the supply chain, the shared constraint is the story: when both camps say they are supply-limited, TSMC's advanced packaging and memory makers' HBM allocation decide the outcome as much as chip design does.