Broadcom Backs Its AI Boom With a $29 Billion Promise
Hock Tan defended Broadcom's AI revenue targets on CNBC and refused to blink at the slowdown debate. Shareholders also carry a quieter promise: a customer-lease backstop worth up to $29 billion if it is ever called.
Vincent Jiangsix-minute read
Hock Tan went on CNBC on September 14 and refused to lower his AI revenue ambitions.1
Broadcom's−10.04% — Broadcom, down 10.04 percent today CEO stood by targets of $115 billion in fiscal 2027 and $230 billion in 2028, despite the debate over slowing frontier-model development. Those are management forecasts. Shareholders also own a less comfortable promise: a customer-lease backstop with maximum potential exposure of roughly $29 billion.1,2
The complication is straightforward. Broadcom earns money helping customers make AI cheaper. It is also helping finance their access to equipment. Shareholders therefore need to assess both the chips' commercial advantage and the financial obligations attached to selling them.
AI dominates the revenue increase
Broadcom's September 2 results reported $29.591 billion of quarterly revenue, including $16.7 billion from AI semiconductors: custom accelerators and networking. The comparable figures a year earlier were $15.952 billion and $5.2 billion. By subtraction, AI supplied about 84% of the revenue increase.4,5
Data
| Value | |
|---|---|
| Q3 FY2025 | $5.2B |
| Q3 FY2026 | $16.7B |
Financing the buyer makes credit judgment part of the product.
Broadcom reported $13.665 billion of company-wide free cash flow after $532 million of capital spending, approximately 46% of sales. Current cash generation is substantial; the lifetime profitability of new financing arrangements remains unproven.4
That distinction is the investment question. A fast-growing supplier can produce excellent current cash flow while accepting obligations whose cost only becomes visible later. The relevant test is whether incremental profit adequately compensates shareholders for those obligations.
The moat is a working relationship
OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T Jalapeño chip explains why Broadcom matters. In its June 24 announcement, OpenAI described designing the architecture around its models and serving systems, with Broadcom contributing silicon implementation and networking. The customer understands the workload; Broadcom helps turn that knowledge into producible hardware.6
This relationship can be difficult to replace because the engineering decisions are connected. Memory movement, software, networking and chip design must work together. A competing component is insufficient if adopting it delays the complete system. The inference: Broadcom's strongest protection comes from its role in repeated development cycles and successful deployments, not any single component.
OpenAI's August 25 tests reported better throughput per watt across three public models. Yet production qualification and software preparation were continuing, with initial deployment planned by year-end. Working silicon is meaningful evidence; profitable operation at scale remains a separate milestone.7
Customers also preserve alternatives. Marvell's−9.26% — Marvell, down 9.26 percent today August 19 filing describes an expanded Google partnership covering custom chips around the TPU ecosystem, including inference accelerators and controllers. It does not establish that Broadcom lost an existing design. It does establish that customers can cultivate additional suppliers while expanding custom computing.8
Financing brings customer risk back to Broadcom
In June, Apollo announced a $35 billion initial financing solution for Broadcom's AI XPV platform, alongside Blackstone and banks. The financing supports more than one gigawatt of Anthropic$350B — Anthropic, private, latest valuation $350B capacity through a multiyear draw schedule. That is financing capacity, not $35 billion of Broadcom revenue.9
Broadcom's June filing disclosed support for five-year customer leases. Exposure rises as racks deploy and falls as lease payments arrive. The September filing specifies approximately $29 billion of maximum potential liability after all racks deploy, undiscounted. At August 2, the backstop's fair value was immaterial and no payments had been made.3,2
Data
| Value | |
|---|---|
| Max backstop liability | $29B |
| Trailing 4Q free cash flow | $39.4B |
Following default, exposure depends on the shortfall between 85% of outstanding backstopped lease amounts and rack-sale proceeds. Other remedies include assuming leases and conditional fixed-price resale. The maximum is neither current debt nor an expected loss.2
The economic concern is that repayment stress and weaker equipment values could arrive together. If an AI customer's economics deteriorate, other buyers might also demand lower prices for the same generation of hardware. That would weaken a source of recovery precisely when it matters. This is a risk mechanism, not evidence that a default is approaching.
Google's April 22 announcement makes the tension concrete. It claims its TPU 8i offers up to 80% better inference performance per dollar than Ironwood. Faster improvement can stimulate demand while pressuring older equipment's rental economics. Google's claim does not establish a resale price or prove that older chips become unusable.10
The strongest bull case survives slower training
Tan's September 14 defense focused on inference: using trained models to complete work. Slowing the release of more capable frontier models would not automatically stop businesses from serving existing ones. More usage can require more capacity even without an equally rapid training race.1
There is commercial evidence behind that argument. Anthropic's May 28 financing announcement said its annualized revenue run rate had exceeded $47 billion earlier that month and disclosed a $65 billion funding round. Those are company claims, and run rate is neither realized annual revenue nor profit. They nevertheless show a customer with substantial sales and access to capital.11
Amazon−2.75% — Amazon, down 2.75 percent today separately said in April that more than 100,000 customers were running Claude on AWS. It also announced additional Trainium capacity for Anthropic. Demand for useful AI can therefore be real while spending is divided across competing hardware platforms.12
Financing support is becoming part of that competition. AMD's−10.57% — AMD, down 10.57 percent today July agreement paired planned Anthropic deployments with up to $5 billion of future equity investment. Nvidia's−10.86% — Nvidia, down 10.86 percent today March partnership with Marvell also sought a role in custom-chip systems. Broadcom competes for the customer's infrastructure budget across engineering, commercial terms and capital support.13,14
The bull case is credible: financing removes a bottleneck, deployments grow, and customers repay from expanding businesses. It does not eliminate the need to price the risk Broadcom retains.
What $344.72 already requires
AVGO closed September 14 at $344.72. Using approximately 4.774 billion shares outstanding on August 28 implies an equity value of $1.646 trillion. The earlier share-count date makes this a valuation snapshot, not an exact current market capitalization.15,2
Trailing four-quarter free cash flow was $39.403 billion, putting that valuation at roughly 42 times actual cash generation.16,2
Annualizing the latest quarter's free cash flow gives $54.66 billion. The estimated equity value is about 30 times that pace, equivalent to a 3.3% cash-flow yield. This is neither trailing annual cash flow nor a forecast: it assumes four quarters matching the latest result. It also leaves stock compensation and any future backstop payments outside an owner-cost adjustment.4
A simple sensitivity shows why growth alone is insufficient. If annual cash flow rose 25% from that pace, but investors paid 25 times cash flow, equity value would increase only about 4%, before dividends and assuming unchanged shares. These are illustrative assumptions, not a price target. A lower valuation multiple can absorb much of the operating improvement.
The bar before treating the revenue targets as shareholder returns should be durable cash generation per share. The business has earned credibility; the valuation still requires continued execution.
What Tan needs to show next
The next disclosures should connect chip shipments to customer deployments, cash receipts and retained financing exposure: Jalapeño's production rollout, the amount of backstop actually outstanding as racks deploy, and whether future transactions require similar or greater support.
Cash per share deserves the same attention as revenue. Broadcom can meet ambitious sales targets while shareholders receive a less attractive bargain if financing exposure grows faster than the profits compensating for it.
Tan has defended the demand forecast. The next test is how much risk Broadcom must carry to fulfill it.
Get Chips by email, free
How AI demand is reshaping chipmakers, memory and the fabs behind them. In our daily email, every weekday.





Reader comments
0 comments