The Tech Sales Newsletter #95: The golden age of silicon
NVIDIA released their Q1'26 earnings results, which was the first full quarter under the new volatile world order of tariffs and rapid AI innovation. After a significant amount of FUD being spread over the quarter, their performance and long-term outlook remains…outstanding.
The key takeaway
For tech sales:
Source: RepVue
For investors: $1000 remains FUD (not financial advice—chalk it up to the AI bull in me).
The power of silicon
Source: NVIDIA company presentation May’25
Colette Kress: We delivered another strong quarter, with revenue of $44 billion, up 69% year over year, exceeding our outlook in what proved to be a challenging operating environment. Data center revenue of $39 billion grew 73% year on year. The AI workloads have transitioned strongly to inference, and AI factory build-outs are driving significant revenue. Our customers' commitments are firm. On April 9, the US government issued new export controls on H20, our data center GPU designed specifically for the China market.
Blackwell contributed nearly 70% of data center compute revenue in the quarter, with a transition from Hopper nearly complete. The introduction of GB200 NBL was a fundamental architectural change to enable data center scale workloads and to achieve the lowest cost per inference token. These systems are complex to build, we have seen a significant improvement in manufacturing yields, and rack shipments are moving to strong rates to end customers. GB200 and VO racks are now generally available for model builders, enterprises, and sovereign customers to develop and deploy AI.
On average, major hyperscalers are each deploying nearly 1,000 NBL72 racks, or 72,000 Blackwell GPUs per week, and are on track to further ramp output this quarter. Microsoft, for example, has already deployed tens of thousands of Blackwell GPUs and is expected to ramp to hundreds of thousands of GB200s with OpenAI as one of its key customers. Key learnings from the GB200 ramp will allow for a smooth transition to the next phase of our product roadmap, Blackwell Ultra. Sampling of GB300 systems began earlier this month at the major CSPs, and we expect production shipments to commence later this quarter.
NVIDIA GPU – Design-to-Delivery Timeline
| Stage | Owner | What happens | Typical elapsed time* |
|---|---|---|---|
| 1. Architecture & RTL | NVIDIA | NVIDIA designs new GPU cores, memory fabric, and power domains. Engineering teams write RTL code and firmware, followed by extensive pre-silicon verification and simulation. | 12–18 mo |
| 2. Physical design & tape-out | NVIDIA | NVIDIA completes floor-planning, timing closure, and power/thermal sign-off. The design is finalized into GDS II files ready for manufacturing ("tape-out"). Multiple risk-production chips are created to validate yield. | 3–6 mo |
| 3. Mask set fabrication | Mask Partners | Specialized mask manufacturers (Toppan, Dai Nippon, etc.) convert NVIDIA's GDS files into ~100–150 photomasks using e-beam writers. At 3nm, mask sets can cost >$15M. | 4–6 wk |
| 4. Wafer fabrication | TSMC | TSMC patterns EUV & DUV layers on 300mm wafers in their advanced fabs. Inline test structures provide early yield data. TSMC manages the complex N3/N4 process flows. | 14–16 wk per lot |
| 5. Backend processing | TSMC | TSMC performs bump attachment, wafer-level testing, and die singulation. Known-good dies are selected while defective dies are marked or mapped out. | 1–2 wk |
| 6. Advanced packaging | TSMC | TSMC's CoWoS-L facility embeds GPU dies on silicon interposers, attaches HBM memory stacks, and completes under-fill and molding. NVIDIA has pre-booked ~70% of TSMC's 2025 CoWoS capacity. | 4–8 wk |
| 7. Final test & binning | NVIDIA | NVIDIA performs power-on testing, functional verification, and speed binning. Firmware is flashed and chips are classified into different SKU tiers based on performance. | 1–2 wk |
| 8. System integration | ODM Partners | ODM partners (or NVIDIA's new U.S. facilities in Texas & Arizona) solder GPUs to HGX/PCIe boards, wire chassis, and perform system-level testing for Blackwell-class systems. | 4–6 wk |
| 9. Deployment & bring-up | End Customer | Customers receive rack deliveries, validate power & cooling infrastructure, run cluster-wide burn-in tests, and qualify CUDA/driver stacks. Hyperscalers automate most processes. | 2–6 wk |
Every new architecture release is coming every 12-18 months and offers significant jumps in efficiency. The biggest priority right now for semiconductors is to keep offering higher performance within the same physical space (ideally also within the same energy envelope) due to constraints on the data center end. There was a lot of FUD in the last months on how Blackwell adoption will be slow and that for the hyperscalers it's smarter to "wait" for the new generation, which never made any sense because these investments are made on a multi-year basis. Nobody is going to mess with this timeline; otherwise you lose your seat at the table.
This week on “How to sell AI” I go further into the hardware part of AI by seeing the actual production process in action and how it relates to the progression we are seeing in model performance.
The definitive tech sales guide to selling AI (LLMs and Enterprise-grade Machine Learning).
Source: NVIDIA company presentation May’25
Colette Kress: GB300 will leverage the same architecture, same physical footprint, and the same electrical and mechanical specifications as GB200. The GB300 drop-in design will allow CSPs to seamlessly transition their systems and manufacturing used for GB200 while maintaining high yields. B300 GPUs, with 50% more HBM, will deliver another 50% increase in dense FP4 inference compute performance compared to the B200. We remain committed to our annual product cadence with our roadmap extending through 2028, tightly aligned with the multiple-year planning cycles of our customers.
We are witnessing a sharp jump in inference demand. OpenAI, Microsoft, and Google are seeing a step function leap in token generation. Microsoft processed over 100 trillion tokens in Q1, a fivefold increase on a year-over-year basis. This exponential growth in Azure OpenAI is representative of strong demand for Azure AI Foundry, as well as other AI services across Microsoft's platform. Inference serving startups are now serving models using B200, tripling their token generation rate and corresponding revenues for high-value reasoning models such as Deepseeker One. As reported by Artificial Artificial Analysis, NVIDIA Corporation's Blackwell NBL72 turbocharges AI inference throughput by 30x for the new reasoning models sweeping the industry.
Developer engagements increased with adoption ranging from LLM providers such as Perplexity to financial service institutions such as Capital One, who reduced agentic chatbot latency by 5x with Dynamo. In the latest MLPerf inference results, we submitted our first results using GB200 NBL72, delivering up to 30x higher inference throughput compared to our 8 GPU H200 submission on the challenging Llama 3.1 benchmark. This feat was achieved through a combination of tripling the performance per GPU as well as 9x more GPUs, all connected on a single NVLink domain. And while Blackwell is still early in its life cycle, software optimizations have already improved its performance by 1.5x in the last month alone.
We expect to continue improving the performance of Blackwell through its operational life as we have done with Hopper and Ampro. For example, we increased the inference performance of Hopper by four times over two years. This is the benefit of NVIDIA Corporation's programmable CUDA architecture and rich ecosystem. The pace and scale of AI factory deployments are accelerating, with nearly 100 NVIDIA Corporation-powered AI factories in flight this quarter, a twofold increase year over year, with the average number of GPUs powering each factory also doubling in the same period.
There is a clear step-up in inference demand in combination with a very hard push on launching new models from the leading providers. This results in every single GPU going on the market being bought and deployed, no exceptions.
Colette Kress: More AI factory projects are starting across industries and geographies. NVIDIA Corporation's full-stack architecture is underpinning AI factory deployments as industry leaders like AT&T, BYD, Capital One, Foxconn, MediaTek, and Telenor are strategically vital sovereign clouds like those recently announced in Saudi Arabia, Taiwan, and the UAE. We have a line of sight to projects requiring tens of gigawatts of NVIDIA Corporation AI infrastructure in the not-too-distant future. The transition from generative to agentic AI, AI capable of perceiving, reasoning, planning, and acting, will transform every industry, every company, and country.
We envision AI agents as a new digital workforce capable of handling tasks ranging from customer service to complex decision-making processes. We introduced the Llama NemoTron family of open reasoning models designed to supercharge agentic AI platforms for enterprises. Built on the Llama architecture, these models are available as NIMS or NVIDIA Corporation inference microservices with multiple sizes to meet diverse deployment needs. Our post-training enhancements have yielded a 20% accuracy boost and a 5x increase in inference speed.
Leading platform companies including Accenture, Cadence, Deloitte, and Microsoft are transforming work with our reasoning models. NVIDIA Corporation's Nexmo microservices are generally available across industries and are being leveraged by leading enterprises to build, optimize, and scale AI applications. With Nexmo, Cisco increased model accuracy by 40% and improved response time by 10x in its code assistant. Nasdaq realized a 30% improvement in accuracy and response time in its AI platform's search capabilities.
Shell's custom LLM achieved a 30% increase in accuracy when trained with NVIDIA Corporation's Nexmo. Nemo's parallelism techniques accelerated model training time by 20% when compared to other frameworks. We also announced a partnership with Yam Brands, the world's largest restaurant company, to bring NVIDIA Corporation AI to 500 of its restaurants this year, expanding to 61,000 restaurants over time to streamline order taking, optimize operations, and enhance service across its restaurants. For AI-powered cybersecurity, leading companies like Checkpoint, Cloudstrike, and Palo Alto Networks are using NVIDIA Corporation's AI security and software stack to build, optimize, and secure agentic workflows, with CloudStrike realizing 2x faster detection, triage with 50% less compute cost.
Source: NVIDIA company presentation May’25
What most still don't understand is that the biggest players are collaborating directly with the ML experts from NVIDIA where possible. There is a significant edge for every player that is big enough and able to engage with the right SMEs behind CUDA. There is a practical reason for it - having more compute doesn't mean that you'll get the expected performance improvement; it requires a much tighter and technically complex approach to software optimization coupled with a deep understanding of what the hardware is capable of.
Colette Kress: Introduced Spectrum X and Quantum X silicon photonics switches, featuring the world's most advanced copacage optics. These platforms will enable next-level AI factory scaling to millions of GPUs through the increasingly power efficiency by 3.5x and network resiliency by 10x while accelerating customer time to market by 1.3x. Transitioning to a quick summary of our revenue by geography. China, as a percentage of our data center revenue, is slightly below our expectations and down sequentially due to H20 export licensing controls.
For Q2, we expect a meaningful decrease in China data center revenue. As a reminder, while Singapore represented nearly 20% of our Q1 build revenue as many of our large customers use Singapore for centralized invoicing, our products are almost always shelved elsewhere. Note that over 99% of H100, H200, and Blackwell data center compute revenue billed to Singapore was for orders from US-based customers.”
It’s important to put into perspective that the record quarter is occurring in the backdrop of a very significant escalation of tensions between the USA and China. While there were already significant limitations on what could be produced and exported - even that compute was essentially blocked.
Source: NVIDIA company presentation May’25
Colette Kress: We have deepened Omniverse's integration and adoption into some of the world's leading software companies, including SAP and Schneider Electric. New Omniverse blueprints such as MEGA for at-scale robotic fleet management are being leveraged by Keyon Group, Pegatron, Accenture, and other leading companies to enhance industrial operations. At Computex, we showcased Omniverse's great traction with technology manufacturing leaders, including TSMC, Quanta, Foxconn, and Pegatron.
Using Omniverse, TSMC saves months in work by designing fabs virtually. Foxconn accelerates thermal simulations by 150x. Pegatron reduced assembly line defect rates by 67%. Lastly, with our automotive group, revenue was $567 million, down 1% sequentially, but up 72% year on year. Year-on-year growth was driven by the ramp of self-driving across a number of customers and robust end demand for NEVs. We are partnering with GM to build the next-gen vehicles, factories, and robots using NVIDIA Corporation's AI simulation and accelerated computing.
We are now in production with our full-stack solution for Mercedes Benz, starting with the new CLA, hitting roads in the next few months. We announced Isaac Groot, N1, the world's first open fully customizable foundation model for humanoid robots, enabling generalized reasoning and skill development. We also launched new open NVIDIA Corporation Cosmo World Foundation models. Leading companies include OneX, Agility Robots, Figueroa, Uber, and Wabi.
We've begun integrating Cosmos into their operations for synthetic data generation. While agility robotics, Boston Dynamics, and XPEN robotics are harnessing Isaac simulation to advance their humanoid efforts. GE Healthcare is using the new NVIDIA Corporation Isaac platform for healthcare simulation built on NVIDIA Corporation Omniverse and using NVIDIA Corporation Cosmos. The platform speeds development of robotic imaging and surgery systems.
The era of robotics is here. Billions of robots, hundreds of millions of autonomous vehicles, and hundreds of thousands of robotic factories and warehouses will be developed.
I've previously covered that outside of AGI, the next most exciting leap in technology will be related to physical AI, essentially embodying models through a variety of form factors. The self-driving car is the most imminent one, but a significant focus is now shifting towards robotics, with Tesla positioning itself as the most likely to deploy factory robots at scale (Optimus). The other players will have to either buy technology from Tesla (which is another significant bullish factor for them) or go to somebody else. The somebody else, of course, is NVIDIA.
Source: NVIDIA company presentation May’25
Now the opening statement in the earnings call is typically delivered by Jensen, but in this case he let the CFO present performance. Instead, he focused on a pointed topic, to say the least:
Jensen Huang: Thanks, Colette. We've had a busy and productive year. Let me share my perspective on some topics we're frequently asked. On export control, China is one of the world's largest AI markets and a springboard to global success. With half of the world's AI researchers based there, the platform that wins China is positioned to lead globally. Today, however, the $50 billion China market is effectively closed to US industry.
The H20 export ban ended our Hopper data center business in China. We cannot reduce Hopper further to comply. As a result, we are taking a multibillion-dollar write-off on inventory that cannot be sold or repurposed. We are exploring limited ways to compete, but Hopper is no longer an option. China's AI moves on with or without US chips. It has to compute to train and deploy advanced models. The question is not whether China will have AI, it already does.
The question is whether one of the world's largest AI markets will run on American platforms. Shielding Chinese chipmakers from US competition only strengthens them abroad and weakens America's position. Export restrictions have spurred China's innovation and scale. The AI race is not just about chips. It's about which stack the world runs on. As that stack grows to include 6G and quantum, US global infrastructure leadership is at stake.
The US has based its policy on the assumption that China cannot make AI chips. That assumption was always questionable and now it's clearly wrong. China has enormous manufacturing capability. In the end, the platform that wins the AI developers wins AI. Export controls should strengthen US platforms, not drive half of the world's AI talent to rivals.
Released freely, they've gained traction across the US, Europe, and beyond. DeepSeek, like ChatGPT, introduced reasoning AI that produces better answers the longer it thinks. Reasoning AI enables step-by-step problem-solving, planning, and tool use, turning models into intelligent agents. Reasoning is compute-intensive and requires hundreds to thousands more tokens per task than previous one-shot inference.
Reasoning models are driving a step function surge in inference demand. AI scaling laws remain firmly intact not only for training but now inference too, requiring massive scale compute. So DeepSeq also underscores the strategic value of open-source AI. When popular models are trained and optimized on US platforms, it drives usage, feedback, and continuous improvement, reinforcing American leadership across the stack.
US platforms must remain the preferred platform for open-source AI. That means supporting collaboration with top developers globally, including in China. America wins when models like DeepSeek and Qwen run best on American infrastructure. Regarding onshore manufacturing, President Trump has outlined a bold vision to reshore advanced manufacturing, create jobs, and strengthen national security. Future plants will be highly computerized and roboticized.
We share this vision. TSMC is building six fabs and two advanced packaging plants in Arizona to make chips for NVIDIA Corporation. Process qualification is underway, with volume production expected by year-end. Spill and Amcor are also investing in Arizona, constructing packaging, assembly, and test facilities. In Houston, we're partnering with Foxconn to construct a million-square-foot factory to build AI supercomputers.
Wistron is building a similar plant in Fort Worth, Texas. To encourage and support these investments, we've made substantial long-term purchase commitments, a deep investment in America's AI manufacturing future. Our goal from chip to supercomputer, built in America, within a year. Each GB200 MBLink72 rack contains 1.2 million components and weighs nearly two tons. No one has produced supercomputers on this scale.
Our partners are doing an extraordinary job. On AI diffusion rule, President Trump rescinded the AI diffusion rule, calling it counterproductive, and proposed a new policy to promote US AI tech with trusted partners. On his Middle East tour, he announced historic investments. I was honored to join him in announcing a 500-megawatt AI infrastructure project in Saudi Arabia and a 5-gigawatt AI campus in the UAE.
President Trump wants US tech to lead. The deals he announced are wins for America, creating jobs, advancing infrastructure, generating tax revenue, and reducing the US trade deficit. The US will always be NVIDIA Corporation's largest market and home to the largest installed base of our infrastructure. Every nation now sees AI as core to the next industrial revolution, a new industry that produces intelligence and essential infrastructure for every economy.
Countries are racing to build national AI platforms to elevate their digital capabilities. At Computex, we announced Taiwan's first AI factory in partnership with Foxconn and the Taiwan government. Last week, I was in Sweden to launch its first national AI infrastructure. Japan, Korea, India, Canada, France, the UK, Germany, Italy, Spain, and more are now building national AI factories to empower startups, industries, and societies.
Sovereign AI is a new growth engine for NVIDIA Corporation.
Now, the point of view presented here puts Jensen in the "positive-sum" camp, basically the belief that AI is to the benefit of everybody globally and should be powered by CUDA and NVIDIA hardware.
This is different from what players such as Palantir, Anduril or the Trump administration see. From their perspective, China cannot be trusted with AGI and it's critical both for the US to be the only one in possession of the technology, as well as to limit the access of untrustworthy partners to the outcomes from it. This is the "zero-sum" approach to AI.
The recent book "Apple in China: The Capture of the World's Greatest Company" had an interesting reframing of what working with China looks like, essentially presenting the idea that Apple ended up subsidizing the rise of Huawei and Xiaomi by training hundreds of thousands of workers in China and building a large network of partners with technical capability.
Since depending on a single customer doesn't make sense for any of them, all of Apple's innovation would typically trickle down between 6 to 12 months later to every other tech company in China. It would be illogical not to assume that the same thing has happened with every other company that outsources high-tech manufacturing in China.
There is a strong argument to be made that DeepSeek, while introducing several new ideas to AI, appears to have been extensively trained by breaking the terms and conditions of OpenAI. "Trickle-down" technology, but this time without a mutual understanding in the background.
Jensen's positioning here is that it's futile to resist this process and instead NVIDIA should drive the hardware adoption of AI in China. Keeping in mind what's at stake (as nicely visualized with the AI 2027 whitepaper that I previously covered), and how Apple's involvement in China has played out over the last 10 years, it's difficult to argue why freely selling GPUs at a large scale is somehow the safer choice. Even if we abstract ourselves from the hypothetical of who gets to AGI first, just the practical implications of limited compute being produced at a time of high demand means that Chinese companies will acquire capacity that could have gone toward Western model training or inference.
The logical consequence of this, of course, is that compute capacity will still be procured in other ways, mostly through black market channels. The ultimate risk, of course, is China seizing Taiwan, and by extension the majority of TSMC capacity (hence his helpful commentary around how partners are ramping up GPU production within US territory).
Joe Moore: You guys have talked about this scaling up of inference around reasoning models for, you know, at least a year now, and we've really seen that come to fruition. As you talked about, we've heard it from your customers. Can you give us a sense for how much of that demand you know, you're able to serve and give us a sense for maybe how big the inference business is for you guys and, you know, do we need full-on NBL72 Rackscale solutions for reasoning inference going forward?
Jensen Huang: Well, we would like to serve all of it. And I think we're on track to serve most of it. Grace Blackwell and VLINK72 is the ideal engine today, the ideal computer thinking machine, if you will, for reasoning AI. There's a couple of reasons for that. The first reason is that the token generation amount, the number of tokens reasoning goes through, is a hundred, a thousand times more than a one-shot chatbot.
You know, it's essentially thinking to itself, breaking down a problem step by step. It might be planning multiple paths to an answer. It could be using tools, reading PDFs, reading web pages, watching videos, and then producing a result, an answer. The longer it thinks, the better the answer, the smarter the answer is. And so what we would like to do and the reason why Grace Blackwell was designed to give such a giant step up in inference performance is so that you could do all this and still get a response as quickly as possible.
Compared to Hopper, Grace Blackwell is some forty times higher speed and throughput. And so this is going to be a huge benefit in driving down the cost while improving the quality of response with excellent quality of service at the same time. So that's the fundamental reason. That was a core driving reason for Grace Blackwell MBLink72. Of course, in order to do that, we had to reinvent literally redesign the entire way that these supercomputers are built.
But now we're in full production. It's going to be exciting. It's going to be incredibly exciting.
Reasoning models have significantly increased the need for inference compute. Reasoning models that operate with a large scope of independence (agents) will only accelerate this process.
Source: NVIDIA company presentation May’25
Jensen Huang I would say compared to the beginning of the year, compared to GTC time frame, there are four positive surprises. The first positive surprise is the step function demand increase of reasoning AI. I think it is fairly clear now that AI is going through exponential growth. And reasoning AI really busted through. Concerns about hallucination or its ability to really solve problems, and I think a lot of people are crossing that barrier and realizing how incredibly effective agentic AI is.
And reasoning AI is. So number one is inference reasoning. And the exponential growth there. Demand growth. The second one you mentioned AI diffusion. It's really terrific to see that the AI diffusion role was rescinded. President Trump wants America to win. And he also realizes that we're not the only country in the race. And he wants the United States to win. And recognizes that we have to get the American stack out to the world and have the world build on top of American stacks instead of alternatives.
And so AI diffusion happened. The rescinding of it happened at almost precisely the time that the countries around the world are awakening to the importance of AI as an infrastructure, not just as a technology of great curiosity and great importance, but infrastructure for their industries and startups and society. Just as they had to build out infrastructure for electricity and the Internet, you've got to build out infrastructure for AI. I think that's an awakening.
And that creates a lot of opportunity. The third is enterprise AI. Agents work. And agents are doing these agents are really quite successful. Much more than generative AI, agentic AI, is game-changing. You know, agents can understand ambiguous and rather implicit instructions and are able to problem solve and use tools and have memory and so on. And so I think enterprise AI is ready to take off, and it's taken us a few years to build a computing system that is able to integrate, run enterprise AI stacks, run enterprise IT stacks, but add AI to it.
And this is the RTX Pro enterprise server that we announced at Computex just last week. And just about every major IT company has joined us, and I'm super excited about that. And so computing is one part of it. But remember, enterprise IT is really three pillars. It's compute, storage, and networking. And we've now put all three of them together finally, and we're going to market with that. And then lastly, industrial AI.
Remember, one of the implications of the world reordering, if you will, is regions onshoring manufacturing and building plants everywhere. In addition to AI factories, of course, there are new electronics manufacturing, chip manufacturing, being built around the world. And all of these new plants and these new factories are creating exactly the right time when Omniverse and AI and all the work that we're doing with robotics is emerging. And so this fourth pillar is quite important.
Every factory will have an AI factory associated with it. And in order to create these physical AI systems, you really have to train a vast amount of data. So back to more data, more training, more AIs to be created, more computers. And so these four drivers are really kicking into turbocharge.
“AI Diffusion” here means the spread of the technology across a variety of business and public sectors.
Source: NVIDIA company presentation May’25
Jensen Huang: This is the start of a powerful new wave of growth. Grace Blackwell is in full production. We're off to the races. We now have multiple significant growth engines. Inference, once the light of workload, is surging with revenue-generating AI services. AI is growing faster and will be larger than any platform shifts before, including the Internet, mobile, and cloud. Blackwell is built to power the full AI lifecycle, from training frontier models to running complex inference and reasoning agents at scale.
Training demand continues to rise with breakthroughs in post-training and reinforcement learning and synthetic data generation. But inference is exploding. Reasoning AI agents require orders of magnitude more compute. The foundations of our next growth platforms are in place and ready to scale. Sovereign AI, nations are investing in AI infrastructure like they once did for electricity and the Internet.
Enterprise AI must be deployable on-prem and integrated with existing IT. Our RTX Pro, DGX Spark, and DGX Station Enterprise AI systems are ready to modernize the $500 billion IT infrastructure on-prem or in the cloud. Every major IT provider is partnering with us. Industrial AI, from training to digital twin simulation to deployment, NVIDIA Corporation's Omniverse and Isaac Groot are powering next-generation factories and humanoid robotic systems worldwide.
The age of AI is here. From AI infrastructures, inference at scale, sovereign AI, enterprise AI, and industrial AI, NVIDIA Corporation is ready.
It’s the golden age of silicon.
Do you want to be a part of it or watch from the sidelines?
