Broadcom's AI Sales Could Double in Two Quarters • The Weighted Average

Broadcom’s September 2 results put AI semiconductor sales at $16.7 billion and forecast $21.7 billion for the following quarter. That Q4 guidance would put the business approximately 100.9% above its Q2 revenue in just two quarters—a reason for infrastructure buyers to secure delivery terms, not assume the expanding supply will be theirs.

Reconstructed on September 7, 2026, from records available by September 3, 2026.

The growth is real; your allocation is not

The most consequential distinction in an earnings call is sometimes between two versions of the word “secured.” A supplier can secure the manufacturing inputs required for its forecast. That does not secure a particular customer’s access to the resulting compute. Broadcom’s report makes a strong case that custom accelerators and AI networking are becoming a larger business. It does not turn a cloud provider’s sales proposal into a guaranteed delivery schedule.

The scale is no longer marginal. Broadcom reported $29.591 billion in total fiscal Q3 revenue, with $20.839 billion from semiconductor solutions and $8.752 billion from infrastructure software in its financial release. The AI semiconductor figure includes custom accelerators and networking. It is not a disclosure of accelerator sales alone, still less a measurement of the compute available to an ordinary enterprise buyer.

That classification matters when comparing suppliers. A custom chip and the network connecting it to other chips form a system, but revenue from the system is not the same as useful inference capacity. Buyers need the model, software environment, interconnect, availability, and operating terms attached to a particular offering. A supplier’s aggregate revenue growth cannot establish its cost per accepted task.

Broadcom’s larger forecast is nevertheless an important planning signal. Reuters’ September 2 report says the company raised its fiscal 2027 AI semiconductor outlook to about $115 billion, from a prior forecast above $100 billion, and expects roughly $230 billion in fiscal 2028. Those are management expectations about future sales, not completed deployments. They suggest a substantial alternative supply chain is forming around custom silicon, while leaving its eventual distribution open.

The immediate operator decision is therefore narrower than “switch away from GPUs.” Cloud procurement teams should ask whether a proposed accelerator service has a credible capacity path and a workload-specific reason to use it. Engineering teams should preserve enough portability to test that service without committing their entire application to its assumptions. A growing alternative is useful leverage only if the buyer can actually run something on it.

This extends, rather than repeats, the archive’s examination of Anthropic’s different costs for rented and owned capacity and the power implications of OpenAI’s Jalapeño accelerator. Those stories concern particular relationships and machines. Broadcom’s new results concern the scale and delivery credibility of the supplier portfolio behind them.

Two quarters turn a workhorse into a constraint

The original calculation joins two separate reporting dates. Broadcom’s June 3 fiscal Q2 release reported $10.8 billion in AI semiconductor revenue. Its September 2 release reports $16.7 billion for Q3 and guides to $21.7 billion for Q4. Subtract $10.8 billion from $21.7 billion and divide the $10.9 billion difference by $10.8 billion: the implied increase is 100.9%, using the company’s rounded figures.

Broadcom's Q4 AI guide is 100.9% above its Q2 sales

AI semiconductor revenue, US$ billions; fiscal 2026

Broadcom Q2 and Q3 earnings releases · June 3 and September 2, 2026

The last bar is a forecast, deliberately distinguished from reported sales. It is not legitimate to describe all three quarters as an accomplished doubling. But the progression changes the procurement conversation even before the final quarter closes. A buyer evaluating next-quarter service needs to ask what part of the supplier’s growth represents usable allocation, rather than repeat an annual capacity aspiration.

The nearer step is substantial on its own. The Q3 release’s Q4 AI guide implies another $5 billion in quarterly AI semiconductor revenue above the reported Q3 level. That is arithmetic on supplier sales, not an estimate of how many racks will become available. Product mix, shipment timing, and networking content can change the relationship between revenue and installed compute.

Management’s explanation locates the constraint in the chain of delivery. The September 2 earnings-call transcript discusses the leading-edge wafers, substrates, memory, sites, and power needed to support its outlook. These are complementary inputs. An accelerator that arrives before its intended site can accept it is inventory, not service capacity; a powered building without the required systems is not an available model endpoint either.

TrendForce’s September 3 account identifies substrates as a particular bottleneck and reports Broadcom’s plan to bring substrate production online in Singapore in fiscal 2027. That is a prospective mitigation, not evidence that today’s constraint has vanished. A capacity proposal should therefore identify the dependency that determines its earliest credible delivery, not merely cite the strongest part of the vendor’s supply chain.

The purchasing lesson is to separate milestones. Manufacturing allocation, shipment, installation, power readiness, software qualification, and customer acceptance should not be interchangeable labels in a contract. Ask the provider to identify which milestone its promised date actually represents. Where the provider cannot disclose its upstream arrangements, seek observable acceptance conditions and remedies rather than attempting to audit a supplier several layers removed.

The distinction becomes sharper when one company sells both hardware and the software surrounding it. SDxCentral’s results coverage notes Broadcom’s interest in VMware Private AI Cloud alongside its semiconductor expansion. Enterprise buyers should evaluate such a platform on its own deployment and governance merits. A prosperous chip business does not automatically make a bundled private-cloud product the economical choice for a particular workload.

Where the supplier thesis can fail

Start with the forecast itself. Broadcom’s Q3 release explicitly warns that actual results may vary materially from guidance and names dependence on contract manufacturing, a limited supplier base, customer-demand timing, and manufacturing capacity among its risks. These are not decorative legal qualifications when the central claim is that supply has been arranged to meet rapidly growing demand. They describe the conditions under which the final bar in the chart fails to arrive.

A second failure mode is mistaking supplier diversification for customer diversification. Reuters identifies major custom-chip customers including Meta, Google, and OpenAI. A business serving a concentrated set of large buyers can grow enormously without creating a competitive service market for everyone else. The operator needs a route to purchase capacity on acceptable terms. Another customer’s deployment plan is not that route.

Competition can also move the value elsewhere. The Reuters account points to increased competition, including Marvell’s recent custom-chip deal with Google. That gives buyers a reason not to treat a supplier’s current position as permanent. It does not establish that switching will be easy: a competing design relationship and an available, qualified cloud service are different stages of the market.

The strongest skeptical case is that no infrastructure switch follows from this earnings report at all. A company consuming managed model APIs may have neither control over the underlying accelerator nor a reason to care which chip executes a request. For that buyer, the relevant outputs remain price, latency, reliability, contractual protection, and task quality. Broadcom’s growth is background information unless it changes one of those terms.

That is a useful limitation, not a weakness to conceal. Microsoft’s new transcription offer can justify a service-level trial without a hardware migration. Meta’s Contributor pricing changes the rights exchanged for inference, not simply the cost of the machine beneath it. Buyers should take the available decision at the layer they actually control.

Even teams that do control infrastructure need a migration denominator. Porting effort, performance qualification, monitoring changes, and any lost flexibility belong in the cost comparison. The sources here do not provide a common benchmark that would justify a universal custom-silicon saving. Quoting one would turn a supply story into an invented efficiency claim. A legitimate business case needs measured results from the intended model and workload on the proposed service.

Evidence that would break the thesis is correspondingly specific: delays in the disclosed supply mitigations, failure to deliver against the near-term guide, or customer capacity that remains inaccessible on commercially usable terms. Evidence that would strengthen it includes on-time service availability, transparent allocation commitments, and repeatable workload economics. A higher long-range revenue target by itself answers none of those questions.

Buy acceptance, not an earnings-slide promise

The practical response is to revise the evidence requested in a procurement process. Treat Broadcom’s outlook as a reason to investigate credible alternatives, not as authority to sign a longer or larger commitment. The release shows commercial momentum; the contract has to specify what a buyer receives, when it receives it, and what happens if the delivered system does not meet its requirements.

For a cloud-capacity purchase, ask for the exact service and region, the availability milestone, the supported software path, and the acceptance test. Distinguish a reservation from usable capacity and an estimated date from a contractual one. Link payment obligations to the milestones the buyer can verify. These are recommendations arising from the supply dependencies described in the call, not claims that any named provider already offers those terms.

For an engineering evaluation, keep the test close to production. Compare the same model or a clearly documented equivalent, the same required output quality, and the same traffic shape. Include retries and failures in the cost denominator. A fast benchmark on a workload the company does not operate offers little negotiating leverage; a reproducible result on an important workload can justify a selective move without demanding a wholesale architectural conversion.

The broader edition reinforces that distinction between a promising input and an accepted outcome. Anthropic’s commerce blueprint still leaves the merchant responsible for valid transactions. WeatherNext 3’s finer forecasts still need a versioned data contract and local validation. More capable infrastructure expands the possible work. It does not perform the buyer’s acceptance testing.

Before the next commitment, the operator checklist is short:

Capacity buyers should request delivery evidence before increasing commitments. Identify the limiting dependency and make the acceptance milestone explicit. The cost at risk is the commitment in the proposed contract, not Broadcom’s aggregate revenue.
Platform teams should qualify a selective alternative, not promise an immediate fleet migration. Measure porting work, operating complexity, and cost per accepted result alongside throughput. Switch only where the measured advantage survives those additions.
Managed-API buyers should negotiate at the service layer. Ask whether supply expansion improves availability, price protection, or service terms. If none changes, an upstream earnings report is not a reason to rebuild the application.
Procurement and finance should revisit the forecast after actual delivery. The Q4 figure remains guidance. Record whether the promised capacity became usable when expected before treating the next forecast as a stronger guarantee.

Broadcom’s results make a persuasive case that AI’s custom-silicon supply chain is becoming much larger. The discipline is to stop that conclusion where the evidence stops. A supplier can secure its supply chain without securing your place in it.