AI Picks and Shovels (Arc 1: Forging the Brain): Broadcom (NASDAQ: AVGO) — The Custom Tailor

Picks and Shovels Series | Arc 1: Forging the Brain | Part 2 of 3

Last week’s piece established the foundry: Taiwan Semiconductor Manufacturing Company (TSM), the plant that physically fabricates the chips that power the AI era. TSMC does not originate the architecture — it manufactures it. Its clients hand over designs; TSMC turns them into working silicon at atomic precision.

Broadcom (NASDAQ: AVGO) is the firm that turns those designs into manufacturable chips — but only for a very specific client list, under a very specific business model that the market spent years misreading, and is now paying serious attention to. It does not invent Google’s TPU architecture from a blank page. It co-designs the thing: translating a hyperscaler’s workload into physical layout, high-speed I/O, packaging and a production programme that can be built at TSMC by the million.

What Broadcom Actually Does

Most coverage of Broadcom frames it as a diversified semiconductor and software company — which is accurate and tells you almost nothing useful about where the value is being created right now.

The more precise framing: Broadcom is the dominant architect of custom AI silicon. It co-designs chips — hardwired, application-specific integrated circuits (ASICs), which the company markets under the XPU label — built to accelerate the specific AI workloads of a specific customer. These are not off-the-shelf processors. They are bespoke, purpose-built accelerators engineered in a multi-year partnership, optimised for that customer’s model architecture, training regime or inference stack, and of limited use at full efficiency to anyone else.

That is the opposite of what Nvidia does.

Merchant Silicon vs. Custom Silicon — The Architecture Divide

Nvidia sells merchant silicon: general-purpose GPUs that any buyer can purchase and run any workload on. The programmability is the product. An H100 running a financial risk model is the same chip as an H100 training a large language model — the architecture is flexible by design.

The flexibility costs something. A general-purpose GPU must allocate transistor budget to programmability — the circuitry that makes it adaptable to diverse tasks. For any single specific workload, that is silicon area and power that is not doing the customer’s actual job.

A Broadcom XPU has far less of that overhead. Transistors are allocated to the operations the customer actually runs. The efficiency gains are not a rounding error. A well-designed ASIC can deliver several times the performance per watt of a general-purpose GPU on the workload it was built for; Broadcom’s own recent characterisation is that custom silicon can match leading GPUs while cutting cost by about half. At the scale hyperscalers are building — tens of thousands of chips in a single cluster, drawing gigawatts — that delta shows up directly in the capex and opex of the world’s largest AI factories.

The trade-off is inflexibility. A Broadcom XPU is a structural bet that the customer’s model architecture will remain stable enough to justify a multi-year co-design programme. If the training approach changes substantially, the roadmap has to change with it. There is no meaningful secondary market for a bespoke chip.

That trade-off is why hyperscalers are doing this at all: they have decided that their architectures are stable enough, at sufficient scale, that the efficiency payoff is worth the engineering commitment.

The Numbers That Made Investors Nervous

On 2 September 2026, Broadcom reported Q3 FY2026 results. AI semiconductor revenue — custom accelerators plus the networking silicon that ties them together — came in at $16.7 billion. That was a 221% year-over-year increase, 54% quarter-on-quarter, and 56% of the entire company. XPUs alone were 73% of that AI figure; XPU shipments were up more than 3.5 times year on year. CEO Hock Tan then guided to $21.7 billion of AI semiconductor revenue in Q4 — a further 236% year-over-year increase — and raised the full-year FY2026 AI target to $58 billion, from $56 billion.

AVGO still fell after the print.

The market’s reaction is now a familiar pattern: when a company’s numbers become extraordinary enough, the extraordinary becomes the floor. The Q3 beat was real. The Q4 company-wide guide of about $34.8 billion sat a sliver below the more aggressive sell-side totals (some consensus sets were near $35.0 billion). In a stock that had already been asked to discount a near-perfect cycle, that gap was enough. The underlying business is not in distress. Management now talks to roughly $115 billion of AI semiconductor revenue in FY2027 and about $230 billion in FY2028, subject to supply and to data centres actually being ready to plug the silicon in.

Visibility is unusual for this industry, even if the oft-quoted $73 billion AI backlog is no longer the live number. That figure was disclosed at the end of FY2025 as AI orders — XPUs, switches, DSPs, lasers — expected to ship over the following 18 months. Much of that book has since converted. The more recent signal is the order flow itself: in Q2 FY2026, Broadcom took more than $30 billion of AI bookings against $10.8 billion of shipments. Demand is not the constraint Tan is warning about. Wafer, packaging, power and building readiness are.

Who‘s Buying — and the Concentration Risk

Broadcom now has six disclosed XPU customers. It does not publish a revenue split, so any percentage column would be an invention. What it has said on the Q3 call, and what is on the public record, is the following:

The concentration risk is still plain, but it has changed shape. Google remains the franchise that built this business — seven generations of TPUs, a contract that now stretches across both compute and networking. If Google brought physical design fully in-house, broke the co-design relationship, or cut capex hard, the near-term trajectory would change. That is the old bear case, and it has not died.

The new one is a three-name problem rather than a one-name problem. Tan’s own 2027–28 map has Anthropic becoming the largest XPU customer, OpenAI the second, and Google still enormous. That is diversification of a sort. It is not safety. Three buyers with correlated AI capex, competing for the same scarce power and the same TSMC CoWoS capacity, can pause together. The same foundry bottleneck we wrote about last week sits under every row of that table.

Broadcom is not unaware of this. The six-customer roster is the answer to the old “Google is the book” critique — but multi-year design cycles do not re-rate overnight, and four of the six programmes are the ones that move the number.

The offset is structural. These are not spot purchases. Hyperscalers take inventory and capacity risk on multi-year wafer commits. Google’s TPU programme is in its seventh and eighth generations. Anthropic is buying the same Ironwood family at gigawatt scale. The switching cost of abandoning a co-design partner mid-programme — lost engineering continuity, accumulated process knowledge, allocated wafers and packaging — is not trivial. Long-term XPU relationships dissolve slowly, when they dissolve at all.

Margin Structure: What Custom Means Financially

Custom silicon comes with a different margin profile than merchant silicon.

Nvidia’s gross margins have consistently run in the mid-to-high 70s, reflecting the pricing power of a dominant, general-purpose platform that customers will pay a premium to access. Broadcom does not break out XPU gross margin on its own. What it does show is the mix effect: non-GAAP consolidated gross margin was about 75% in Q3 and is guided to about 73% in Q4, as XPUs and the high-bandwidth memory inside them become a larger share of the company. Management’s instruction to investors is to watch operating margin instead. That line was 67.9% in Q3 and is guided to about 66% in Q4 — held up by operating leverage, not by software-like chip pricing.

The percentage gap versus Nvidia sounds like a disadvantage. In cash it is not obviously one. Broadcom printed $16.7 billion of AI semiconductor revenue in a single quarter from a line that barely existed three years ago, and $13.7 billion of free cash flow for the company as a whole. Dollar gross profit at this volume compensates for a lower rate — and as design work amortises across more generations and more customers, the rate should not have to do all the work.

The more useful comparison is not Broadcom versus Nvidia on margin percentage. It is Broadcom versus what else a hyperscaler can do with the same silicon budget. On that comparison, XPUs win on power efficiency and total cost of ownership at the scale these buyers actually operate. The margin structure reflects that the value is shared between Broadcom and its customers, not extracted the way a CUDA-locked GPU is.

The Networking Foreshadow

Broadcom is not only a custom-chip company. Its Ethernet switching franchise — Tomahawk for AI scale-out, Tomahawk Ultra for Ethernet scale-up, with Jericho still in the data-centre routing layer — is the connective tissue of modern clusters. Those clusters now need high-throughput, low-latency links between thousands, and eventually hundreds of thousands, of accelerators. AI networking revenue grew more than 2.5 times year on year in Q3; Tomahawk 6 is ramping and Tomahawk 7 has been introduced.

We will return to networking silicon in Arc 2. It is worth noting here: Broadcom’s position in AI infrastructure is not limited to the chip that does the thinking. It also makes the silicon that handles the talking — plus the DSPs, optics and co-packaged optics that sit on the same bill of materials. As clusters grow, the networking layer becomes as critical as the compute layer.

A custom brain is still just a brain — until it can talk to a thousand others.

The Aussie Angle

Here is something Australian investors rarely hear in US-centric coverage: the cloud regions expanding in this country — AWS in Sydney, Google Cloud in Sydney and Melbourne, Azure across multiple cities — sit on the same global silicon stack that Broadcom and TSMC serve. That does not mean a Sydney availability zone is a TPU factory. The giant custom-XPU clusters are still concentrated in a handful of AI campuses, mostly in the United States, where power and land can be assembled at gigawatt scale.

What is local is the dependency. Every major cloud platform an Australian enterprise or agency will use for low-latency or “sovereign-enough” AI still runs on accelerators, CPUs and — more relevant to Broadcom’s everyday footprint here — Ethernet, optics and offload silicon designed offshore. There is no Australian alternative at this layer. Data-sovereignty debates that never mention the chip and the switch are debates about software policy sitting on imported physics.

For an investor holding AVGO through a US equities account, that is indirect exposure to the infrastructure under Australia’s own AI build-out, not only to US hyperscaler capex. It is also a mega-cap with a growing dividend, which is how a lot of SMSF money will justify sitting with a name this volatile.

One practical consideration sits on top of the multiple. AVGO closed at $357.90 on 4 September 2026, off a 52-week high of $495. Trailing earnings put the stock in the mid-40s times; on forward estimates it is closer to twenty times — cheaper than the mid-cycle narrative still circulating, not cheap in any absolute sense. An Australian buyer converting AUD to USD pays a currency overlay on that multiple. If the AUD strengthens against the USD, the AUD-denominated return compresses even if the stock works. If the AUD weakens, the opposite happens. The quality of the business does not change that arithmetic either way.

The Investment Case — In One Line Per Side

Bull case: Broadcom is the clear lead co-design partner for custom AI silicon, with six XPU customers, XPUs already 73% of a $16.7 billion quarterly AI book, a raised map of $58 billion / $115 billion / $230 billion of AI semiconductor revenue across FY2026–28, and a networking franchise that monetises the same clusters a second time.

Bear case: The forward book is still an oligopoly — Google, Anthropic and OpenAI will dominate 2027–28 if Tan’s own gigawatt path is right. Those three share a capex cycle, a foundry and a power grid. Gross margin is being diluted by XPU and HBM mix on purpose. And the stock still trades the guide, not the beat: one light quarter of company-wide revenue is enough to reclassify a 221% AI print as a disappointment.

The honest read: Broadcom is a high-quality business in a high-stakes cycle. Concentration has broadened from one customer to a handful; it has not gone away. Bookings and multi-year wafer commits provide unusual visibility. The networking business is a second structural lever that the pure custom-chip framing misses. Whether the current multiple — mid-40s trailing, nearer 20 times forward, after a one-quarter-plus drawdown from the highs — pays you for that concentration is a question every investor has to answer for themselves.

Coming Next Week — Part 3 of 3

The final piece in Arc 1 visits a company that has built its identity around being the alternative — the challenger to an entrenched incumbent in every major compute cycle it has participated in. It has won those fights before. Whether it wins this one depends on how hyperscalers divide their silicon budget between custom and off-the-shelf compute — and whether this company’s new architecture is a genuine threat or merely a credible enough threat to change the incumbent’s pricing.

Arc 1 closes next week.

Data Sources:

• Broadcom Q3 FY2026 earnings release and earnings call, 2 September 2026 (AI semiconductor revenue $16.7B; Q4 AI guide $21.7B; FY2026 AI $58B; FY2027 ~$115B; FY2028 ~$230B; six XPU customers; XPU mix 73% of AI revenue)
• Broadcom Q4 FY2025 earnings call (origin of the $73B AI backlog figure; ship window described as ~18 months from that date)
• Broadcom Q2 FY2026 earnings commentary (AI bookings of more than $30B against $10.8B shipped)
• OpenAI / Broadcom Jalapeño announcement, 24 June 2026
• Market data as of 4 September 2026 close: AVGO $357.90; 52-week range $289.96–$495.00; trailing P/E mid-40s, forward P/E ~20x depending on the estimate set used

Wall St. Down Under | Australia

Subscribe | wallstdownunder.com.au

Disclaimer: Wall St. Down Under is an independent financial newsletter for informational and educational purposes only. Nothing published here constitutes financial advice, a recommendation to buy or sell any security, or a solicitation of any investment decision. Always do your own research and consult a licensed financial adviser before making investment decisions. Australian investors should consider their own financial situation, objectives, and risk tolerance. Past performance is not indicative of future results.