预计AVGO的AI半导体收入将在2027财年达到约1150亿美元，并在2028财年达到约2300亿美元，基于已锁定的供应和强劲的客户需求。

Welcome to the Broadcom Inc. third quarter 2026 results conference call. Broadcom distributed a press release and financial statements after the market closed detailing our financial performance for the third quarter of fiscal year 2026.

We have had an exceptional quarter, with revenue, operating income and free cash flow all exceeding previous records. The driving force behind this was our AI semiconductor revenue in the third quarter, which grew by 221%. On an annual basis and with a 54% increase on a quarterly basis.

This has resulted in our consolidated revenues reaching 29.6 billion, an increase of 86% year-on-year. Operating income grew at a faster rate of 92% year-on-year, with the operating margin reaching a record high of 68% of revenue, reflecting the strength of operating leverage. Demand in the third quarter was extremely high, and we are still only at the beginning.

Our six XPU customers are accelerating the adoption of custom accelerators, and AI semiconductor revenues have more than tripled year-over-year to $16.7 billion. During the quarter, we supplied Ironwood TPU V7 Version 7 in bulk to both Entropic and Google.

Meanwhile, we have begun production shipments of the next-generation TPU version 8i for Google. This new generation TPU is designed with more memory and bandwidth compared to Ironwood.

As "Ironwood" has been optimized for inference workloads, its performance rivals or even surpasses that of the "Vera Rubin" graphics processing unit.

To further emphasize the major technical challenges of developing these complex accelerators, Broadcom has now started shipping the TPU 8i ahead of MediaTek's release, which actually began earlier.

In the third quarter, we also shipped Jalapeno, the first generation of custom accelerators from OpenAI, which outperforms the Grace Blackwell GPU for inference workloads.

Overall, our shipments of XPUs for the third quarter increased by more than three and a half times year over year, and accounted for 73% of AI revenue during the quarter. Our revenue from AI networks has increased by more than two and a half times year-over-year, and this strong momentum continues into the fourth quarter.

During this quarter, we expect to accelerate Ironwood's shipments to Entropic.

We also expect increased shipments of TPU units version 8i to Google in large sizes. Shipments of "Jalapeno" to "OpenAI" will continue. As for Meta, we anticipate production shipments of their custom NTIA accelerator, which is optimized for inference and recommendations at scale.

In the fourth quarter, we expect revenues from both XPUs and AI networks to triple year-over-year. Together, we expect these deployments to drive our AI revenue in the fourth quarter to $21.7 billion, a 236% year-over-year increase.

Based on this fourth-quarter guidance, I should say that we expect our AI revenue for fiscal year 2026 to be around $58 billion, up 186% year-over-year, which is higher than our previous guidance of $56 billion.

We continue to see exponential growth in demand from our customers for XPU units. We believe that the vast majority of computing demand for AI workloads today stems from this concentrated group developing cutting-edge, leading-edge models, and we expect their need for computing infrastructure to increase even further in 2027 and 2028.

They are all moving towards XPUs to achieve superior performance, cost, and power. Let me now tell you about the journey of each one of them, the journey of each of our customers towards using XPUs on a large scale to power their world-leading models and workloads.

Our connection to Google has never been stronger. This has recently been reinforced by a long-term agreement to develop and supply future generations of TPU modules and AI networks.

Under this agreement, we plan to deliver tens of billions of dollars worth of TPU units annually over the next few years. We expect this increased demand in 2028 and 2029 to be met by successive generations of more sophisticated TPU modules that we are developing today with Google.

Our partnership with Google will continue because we have the strongest intellectual property portfolio in semiconductor design, including industry-leading studies, chip interconnection, advanced HPM and SRAMM integration, and simply superior, advanced packaging.

First and foremost, we have consistently – or rather, I should say consistently – delivered the fastest time to bring TPU modules to market, from the product definition stage to production without the need for modifications.

We believe that these are very deep obstacles that would be difficult for any competitor to overcome. Moving on to Entropic. We're starting with 1 GW of "Iron Wood" that we're deploying in 2026.

We expect Entropic to deploy an additional 5 GW of the TPU version 8i in 2027. And in 2028, we have a clear vision to deliver an additional 10 GW, even as we anticipate Google's business growing with us.

Entropic is on track to become our largest XPU customer in 2027, and will maintain that position in 2028. For OpenAI, Jalapeno is on track for its planned 1.3 GW deployment in 2027.

In collaboration with OpenAI, we are in an advanced stage of developing the next generation of XPUs beyond Jalapeno, which is nearing launch. In 2028, we have a clear vision to deploy over 5 GW of Jalapeño and next-generation XPUs at OpenAI, which will make OpenAI our second-largest XPU customer.

In addition, we are working on developing the third generation of their XPU units in collaboration with OpenAI.

As OpenAI announced last week, the Jalapeno outperforms the Grace Blackwell Ultra in terms of performance per watt, response time, throughput, and power consumption, and is actually comparable to Rubin GPUs in running OpenAI tasks.

The lesson here is that when you develop a custom-designed and optimized chip for your large language model (LLM) tasks, you will outperform any GPU. Such as TPU units. The Jalapeño proves that it is also capable of running other flagship models, and you can do all this at half the cost of a GPU.

Our partnership with Meta to provide multiple generations of MTIA XPUs remains on track. From now until the end of 2027, we will be delivering three generations of MTIA accelerators to Meta Inc.

Across these three generations, we have a clear vision to deploy 3 gigawatts by 2028. Our AI content, as you know, goes beyond just XPUs. We are pioneers in AI networks, and we continue to enhance our superiority in transforming networks for vertical and horizontal expansion.

We were the first to bring the Tomahawk 6 100Tbps switch to market, and we have just produced the Tomahawk 7, the industry's first 200Tbps Ethernet switch for internal expansion.

We continue to lead in every generation of PCI Express conversion. We are now pioneers in advanced digital optical signal processors, and are rapidly expanding our capabilities and market position as leaders in EML, Vixels, and CW laser optical coherence technologies.

In short, we continue to invest heavily to provide the broadest and most advanced portfolio of AI technologies. In fact, our AI network revenues are expected to grow at the same rate as data processing units (XPUs) over the next few years.

Reflecting our excellent progress with this key group of large language model customers, here are our AI semiconductor revenue forecasts. In 2027, we have secured the supply to double AI revenues again to approximately $115 billion.

Our demand actually exceeds these expectations and we will work to improve supply in 2028.

We expect the growth trajectory to continue. We have a clear vision for AI semiconductor revenue growth for fiscal year 2028 to double again to $230 billion.

And here again, we have secured supplies to meet these expectations. These AI revenue guidance through 2028 were provided to illustrate our growth trajectory. Demand for computing remains extremely strong.

As a result, I must say that we are very well on track to surpass 38. Now moving to non-AI semiconductors, revenues reached $4.2 billion, up 5% year-over-year and stable compared to the previous quarter.

Broadband and server storage revenues both increased, which was partially offset by a decline in the wireless sector. In the fourth quarter, we expect non-AI semiconductor revenues to be around $4.3 billion. With another 5% increase compared to the previous quarter.

Speaking of infrastructure software, third-quarter revenue was $8.8 billion, up 29% year-over-year, and we maintained an annualized recurring revenue growth rate of 15% year-over-year.

For the fourth quarter, we expect infrastructure software revenue to stabilize at around $8.7 billion. We have announced VMware's AI private cloud, which gives businesses a secure and cost-effective platform to build and run AI alongside their existing applications.

It combines AI infrastructure, security and compliance, and tools for building and running trusted AI agents, all while protecting enterprise data. VCF also makes it easier for customers to move workloads from the public cloud to the private cloud, where they can gain greater control and significantly improve infrastructure economics.

Enterprise adoption of artificial intelligence actually opens up a new opportunity for our infrastructure software business.

So, in short, for everything related to the quarter, $34.8 billion, an increase of 93% year-over-year. We expect AI revenues in the fourth quarter to reach $21.7 billion, a 236% year-over-year increase, and we expect the operating margin to be around 66% of revenues.

Consolidated revenues reached a record high of $29.6 billion for the quarter, an increase of 86% year over year. Gross profit margin was 75% of revenue in the quarter, down 210 basis points from the previous quarter, as AI semiconductor revenue represented a larger proportion of our total revenue mix. This was better than our guidance of 74%.

Third-quarter operating income reached a record high of $20.1 billion, up 92% from last year. Despite a decline in gross profit margin due to the revenue mix, the operating margin increased by 240 basis points year-on-year to 67.9%. This is thanks to the enormous operational leverage we achieve.

In line with this, non-GAAP earnings per share for the third quarter rose 96% year-over-year to $3.32. Starting with the semiconductor sector, semiconductor solutions revenues reached a record high of $20.8 billion, an increase of 127% year-over-year, and represented 70% of our total revenues.

AI semiconductor revenues accounted for $16.7 billion, or 56% of total revenues, up from 49% in the second quarter.

Our gross profit margin for the semiconductor solutions sector reached approximately 76%. Operating expenses of $1.2 billion reflected investments in research and development, with operating expenses representing 6% of the sector's revenues. Operating margin rose by 61%, an increase of 440 basis points year-on-year, with revenues growing by 127%. This exceeded the growth in operating expenses, which rose by 22% year-on-year.

Moving on to infrastructure software, revenues rose to $8.8 billion, a 29% year-over-year increase, representing 30% of our total revenue. Gross profit margins for infrastructure software reached 94% during the quarter, and operating expenses exceeded $900 million.

The operating margin for software in the third quarter increased by 650 basis points year-on-year to approximately 84%. Turning to the balance sheet, we ended the third quarter with $24 billion in cash compared to $19.6 billion in the previous quarter, an increase of $4.3 billion on a quarterly basis.

We ended the third quarter with $4.5 billion in inventory to support strong demand for semiconductors. Turning to cash flow, free cash flow in the quarter reached a record high of $13.7 billion, representing 46% of revenue. We spent $532 million on capital expenditures during the quarter.

By switching to a capital distribution in the third quarter, we paid shareholders $3.1 billion in cash dividends based on a quarterly dividend of 65 cents per share. In the third quarter, we also repaid $5.6 billion of long-term debt.

After the end of the quarter, we paid an additional $1.5 billion in premium bonds as they matured. The weighted average interest rate and number of years of maturity for our total fixed-rate debt of $59.6 billion are 4% and 7.4 years, respectively.

In June, we established the AI XPV platform in partnership with Apollo and Blackstone to enable more than 20 GW of computing infrastructure for OpenAI and Anthropic by the end of 2028.

We closed the first $35 billion tranche in June to deploy 1 GW of capacity for Anthropic, which is already being implemented. While future chips will include unique features designed to meet the needs of specific labs and investors, our core strategy remains unchanged.

First, we are enabling two of our most important strategic clients, leading AI labs, to bridge the gap between their current cash flow and the large initial investments required for their businesses.

Our XPUs enable sustainable and cost-effective growth at twice the cost of their infrastructure. Secondly, this XPV platform enables customers to attract investments and facilitates implementation where demand is already guaranteed.

Third, we review any strategic financing from a commercial and financial perspective in line with our current capital allocation framework. Through the XPV platform, we partner with specialized external financial partners to independently secure and finance assets rather than providing direct financing ourselves.

Where necessary, we may offer limited residual value guarantees, which are conditional commitments that we consider to be low-risk, supported by a strong profitability trajectory for these coefficients and the sustainable value of the underlying assets.

Turning to the outlook, our forecast for consolidated fourth-quarter revenue is $34.8 billion, a 93% our forecast for consolidated fourth-quarter revenue is $34.8 billion, a 93% year-over-year increase.

We expect semiconductor revenues to reach approximately $26.1 billion, an increase of 136% year-over-year. As part of that, we expect AI-related semiconductor revenues to reach $21.7 billion in the fourth quarter. With an increase exceeding 236% year-on-year.

We expect infrastructure software revenues in the fourth quarter to be around $8.7 billion, an increase of between 24% and 25% year over year. With AI revenue growth accelerating in the fourth quarter, we expect our consolidated gross profit margin to be around 73%, down from 78% a year ago.

As we discussed previously, this increasing mix of XPUs with high memory content reflects a decrease in the consolidated gross profit margin.

However, we expect the operating profit margin in the fourth quarter to be around 66%. This is a stable level compared to last year because our strong revenue growth is driving significant operating savings.

We expect the non-GAAP tax rate for the fourth quarter and fiscal year 2026 to be around 16%, due to the impact of the global minimum tax and the geographical mix of income compared to fiscal year 2025.

We expect the number of non-GAAP diluted shares for the fourth quarter to be around 4.94 billion shares, excluding the impact of potential share buybacks. In the fourth quarter, we expect capital expenditures of $1.4 billion as we invest in increasing semiconductor production capacity.

we expect AI revenues to double again to about $115 billion in fiscal year 2027, and double again in fiscal year 2028 to $230 billion.

We provide these AI revenue projections to illustrate our growth trajectory, and we do not intend to update them on a quarterly basis.

Can you talk about the supplies involved, what the supply bottlenecks are, and what variables might drive this number up if you could resolve them? No matter what I tell you, you keep posting bigger numbers. I find this funny. We are very careful, and to be honest, we try to be discreet. So, we give you our expectations, and yes, the demand...we can ship much larger quantities, but the question that sometimes crosses our minds is: Will the chips we ship be deployed in a timely manner? This is something we always keep in mind when making those predictions.

But our customers certainly want us to ship more. However, we have secured the supplies, and we believe that the number 115 is the correct number. If circumstances change and our ability to secure more supplies changes, we will of course raise expectations. But at this stage, this is our best expectation.

Incidentally, the same thing, and the same way of thinking, applies to 2028 when we gave you a forecast of 230 billion.

We believe this is a genuine demand, based on what is available, what data centers and sites are ready by 2028 in relation to our customers, the size of what we have, and compared to the supply chain we have in advanced chips, substrates, and HBM memory.

So, these are once again carefully structured expectations, and we believe we are capable of achieving them. Can you talk about the substrates, interconnection technologies, or "CoWoS" alternatives? Have you decided to use Singapore's capacity, and how does that fit into the supply picture you have put in place?

Well, we will start operating the Singapore pile factory, by the way. Starting from fiscal year 2027, I believe that will address a major part of our supply bottlenecks. The figure you mentioned for Google, then you said 10 gigawatts for Anthropic. Are they...not number one, are they? The 10 gigawatts are allocated to Anthropic only.

10 gigawatts are dedicated solely to Anthropic. Hawk, your customers are pushing their XPUs and GPUs towards higher performance. And the network's bandwidth expands accordingly, doesn't it? Your customers are now moving from 100 Gbps per route to 200 Gbps per route. From Tomahawk 5 to Tomahawk 6. We've heard that the Tomahawk 6's growth rate has been the fastest among all your conversion families, and we've also heard that you've sold almost your entire Tomahawk 6 inventory for next year. I'm not sure if you can provide us with an update on this.

It's also good to see the team launching its next-generation Tomahawk 7 with a 200GB interface. Leading model customers using XPUs: How many of them use Tomahawk networks for horizontal scaling (scale-out)? And can you also provide us with an update on the adoption of the Tomahawk Ultra platform for vertical scaling (scale-up)?

You are right. Tomahawk 6 has been a tremendous success. As I mentioned, it has already begun to expand horizontally. It is available in both 100GB and 200GB versions. So we actually have two versions of Tomahawk 6. Both are very successful and are deployed in almost all major cloud companies that build XPUs with us, and even those that do not use our units use Tomahawk 6 in both its 100GB and 200GB versions.

With regard to the Tomahawk Ultra, we innovated ahead of the market here by enabling vertical expansion using low-latency Ethernet. We were also surprised by the adoption rate of this device, and we began to see it deployed starting this quarter and in the next fiscal year 2027 in vertical expansion applications.

Just to clarify what Charlie is saying, with Tomahawk Ultra, we are now enabling vertical expansion within a GPU/XPU cluster in a single cabinet, primarily over an Ethernet network for the first time. Because Tomahawk Ultra will operate with the same efficiency as anything that existed before Ethernet, and before the ability to do so over Ethernet, especially with regard to loss and delay.

I just wanted to double-check my calculations. If I add up the gigawatts for 2027 and 2028, I get approximately 10 for 2027, of which about 6 are for my companies Anthropic and OpenAI, and about 20 for 2028, of which about 15 are for my companies Anthropic and OpenAI. I just want to make sure that's correct, and if so, based on my calculations for revenue per gigawatt, I find it to be between 11 and 12 billion per gigawatt. I mean, your competitor mentioned a number close to 40, and I'm wondering: Is that the right level we should be thinking about regarding your content? And how should this be heading in the future, as you mentioned, with more generations of XPU units, which feature higher performance and more memory? What are those numbers trending towards? So, my first question is about the number of gigawatts, and the second is about the content.

Yes, you can calculate the number of gigawatts that we specify. Not everyone, because we focus on being fair, with six clients. Four of them will simply be huge, and we outline and review with you their journey in deploying XPUs. As I said, it's a journey that Google has been undertaking for 10 years, others like OpenAI for two or three years, and others for three years; Each one of them does it differently, and we'd like to take you on that journey, and more importantly, what it means in their deployment of XPUs for the years 26, 27 and 28, so we're putting the numbers around, and you're right, Stacey, if you add up the gigawatts, we're showing the direction they're going.

What we did not mention to you specifically when we reached the final number is the number of gigawatts that will actually be deployed, because deployment is not limited to just providing chips. Before the chips can be supplied, the data center infrastructure must be fully ready for actual production. We are talking here about fiscal years. What we are saying is that when we collect the gigawatts, we are not claiming that over the next two years, 2027 and 2028, 30 gigawatts will enter the production phase and therefore we will ship them. We believe, with a conservative estimate, that the number will be somewhat lower. But we are seeing demand that suggests if they can get everything ready, supply our products, and ship those racks containing the chips, it will reach 30 gigawatts among our six current customers.

The question is, will the entire 30 gigawatts enter the production phase within these two fiscal years? We are giving you an estimate of $115 billion worth of chips for these in fiscal year 27, and another $230 billion in fiscal year 28, which will total, as far as I know, about $350 billion. So, another way to say it is that we believe with a very high degree of confidence. That we will ship $350 billion worth of AI semiconductors to these customers in the next two years. This is the best way to look at it. This does not necessarily mean that the 30 gigawatts must have been fully deployed during that period.

Moving on to your next, more interesting question, right? As I said, when you make an XPU processing unit, it not only performs well, if not better, for the large language model workloads of each of our clients. It is also half the cost, and in fact, less than half the cost. This perfectly illustrates my point of view.

Can you shed more light on the maximum off-budget risk for guarantee agreements? Your last quarter showed that the first tranche had a maximum exposure of approximately 29 billion. Is this roughly the scale level for both Anthropic and OpenAI gigawatts over the next few years? Can you also add more details about the residual value and the risks associated with it?

Listen, we have nothing to announce today regarding residual value guarantees or warranty agreements. Therefore, there is nothing new to add. The figures I mentioned about what we have actually done remain correct. As I mentioned in my prepared notes, we will evaluate any strategic financing on a deal-by-deal basis, and we expect that any financing we undertake in the future will have unique features and will be tailored to suit the needs of the laboratory and the investor. Therefore, I cannot give you a comprehensive overview of the maximum or what each will look like, but we will tell you in due course. We have nothing to announce today.

Amy, I just wanted to clarify any impact on gross profit margins in 2027 and 2028 given this XPU mix and the high cost of memory. Then, you mentioned that Frontier Labs would be your biggest AI customers, even though Google continues to say it is facing supply constraints. So why don't they take more? In many cases, these leading labs rely on land, energy, and infrastructure from their cloud service providers, and sometimes entities affiliated with Nvidia. So to what extent is their use of silicon a choice of their own versus what is dictated by the cloud provider or funding partner? I'm just curious, what gives you the certainty that they'll award Broadcom that particular job in 2027 or 2028 when so much of their funding depends on other cloud service providers who may have a different view when it comes to choosing silicon for that particular facility?

I have two ways of answering that, and two points: Don't forget the growth rate of these startups, and we're talking about six clients and we're creating these financial instruments,

This makes investing in empowering these people economically logical for Broadcom.