Last week, I published Even More Of The Latest Thoughts From American Technology Companies On AI (2026 Q3). In it, I shared commentary in earnings conference calls for the third quarter of 2026, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large.
A few more technology companies I’m watching hosted earnings conference calls for 2026’s third quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:
- 2023 Q1 – here and here
- 2023 Q2 – here and here
- 2023 Q3 – here and here
- 2023 Q4 – here and here
- 2024 Q1 – here and here
- 2024 Q2 – here and here
- 2024 Q3 – here and here
- 2024 Q4 – here, here, and here
- 2025 Q1 – here and here
- 2025 Q2 – here and here
- 2025 Q3 – here, here, and here
- 2025 Q4 – here, here, here, and here
- 2026 Q1 – here, here, here, and here
- 2026 Q2 – here, here, here, and here
With that, here are the latest commentary, in no particular order:
Nu Holdings (NYSE: NU)
Nu Holdings’ management thinks the arrival of AI agents heralds the end of a lot of the profit pools for incumbent US banks
We think that what is happening today with AI and with AI agents, it is the beginning of the end for a lot of those profit pools, because that is just going to bring so much transparency to the market that will be a very, very important catalyst for some of those profit pools to be rearranged in the market.
Nu Holdings’ management has developed version 3 (V3) of its NuFormer foundational AI model for credit; management thinks NuFormer could potentially help Nu Holdings understand new markets faster; management also thinks NuFormer could possibly help Mexicans, Colombians, and Brazilians in the US get access to credit fast
We talked a lot about AI and our nuFormer model about a couple of years ago that we published our first foundation model. I think we were one of the first, if not the first, financial services firm in the world that built its own AI foundation model with some data for credit. We are in V3 of that. One of the interesting potential applications is that that model, the same way your LLM, your favorite LLM, is trained on a language and then helps you learn a new language as you train it with less data, the same way our foundation model from one country or for several countries where we have over 140 million customers can help us faster understand new markets and new populations. This is something that we still kind of need to prove out, but we see good evidence for that.
There’s also other opportunities. So for example, since a lot of our first customers in the U.S., and a meaningful percentage there are Mexican, Colombians, and Brazilians, we can use the credit information that we have on those customers in our core markets to help them get a credit card faster. One of the big pain points you will hear from a lot of these populations is that they are new in the U.S., it might take them years to get a credit card, even though they had a very good credit in their home country. So that is something that we can immediately or fast enough solve given the significant information we have.
NVIDIA (NASDAQ: NVDA)
NVIDIA’s next-generation GPU system, the Vera Rubin, is now the fastest product ramp in the company’s history; Meta Platforms intends to deploy more Vera Rubins, specifically for AI agents; Vera Rubin delivers 10x more tokens per watt over the previous GB300 (Grace Blackwell) platform; management expects software optimisations to help Vera Rubin deliver even better performance over time; the transition from Grace Blackwell to Vera Rubin has been smooth; production racks of Vera Rubin are landing in data centers just 3 months after NVIDIA began production shipments; NVIDIA’s hyperscale customers brought Vera Rubin into production in just 1 month, 2x faster than with Grace Blackwell; management is on track to launch Vera Rubin Ultra in 2027 (FY2028), and Feynman in 2028 (FY2029)
Vera Rubin is off to an extraordinary start. In its first quarter of production shipments, the Rubin platform generated nearly $20 billion in revenue, making it the fastest product ramp in NVIDIA’s history…
…To support this massive scale and future growth, Meta intends to expand its deployment of Vera Rubin hardware, with the goal of maximizing performance and efficiency for its next generation of AI agents…
…On the MLPerf Endpoints benchmark, Vera Rubin delivered 10x more tokens per watt over GB300. With Vera Rubin still early in its life cycle, we expect continuous software optimization to deliver even greater performance over time. The transition from Grace Blackwell to Vera Rubin has been exceptionally smooth. Just three months after we began production shipments in August, production racks are landing in data centers at every major hyperscaler and leading neoclouds. Our hyperscale partners brought Vera Rubin into production in about one month, in half the time required with Grace Blackwell…
…We remain on track to launch Vera Rubin Ultra next year and Feynman in 2028.
Hyperscale revenue grew 23% sequentially in 2026 Q3 (FY2027 Q3), and has accelerated for 6 consecutive quarters; ACIE (AI Clouds, Industrial, and Enterprise) revenue grew 18% sequentially in 2026 Q3 (FY2027 Q3); sovereign AI revenue more than tripled year-on-year in 2026 Q3 (FY2027 Q3); physical AI revenue nearly tripled year-on-year in 2026 Q3 (FY2027 Q3)
Hyperscale revenue of $60 billion grew 23% sequentially, driven by the Vera Rubin ramp and sustained strength in Grace Blackwell. Our hyperscale customers’ cloud revenue growth accelerated for the sixth consecutive quarter…
…AI Clouds, Industrial, and Enterprise, or ACIE, revenue of $47 billion increased 18% sequentially and nearly tripled year-over-year. Sequential growth was driven by Vera Rubin shipments to top leading neocloud customers and the start of our frontier labs building their own data centers…
…Sovereign AI revenue more than tripled year-over-year in Q3…
…Physical AI revenue nearly tripled year-over-year in Q3 and is now on track to exceed $25 billion for the full year.
NVIDIA’s management is seeing strong demand for AI compute from a diverse group of customers, from hyperscalers and frontier labs, to governments
We are addressing strong demand for compute across a diverse and expansive set of end customers, including hyperscalers, frontier labs, AI natives, enterprises, and governments.
NVIDIA’s data center revenue had strong sequential growth of 21% in 2026 Q3 (FY2027 Q3), with all subsegments contributing to the performance
Q3 data center revenue increased 21% quarter-over-quarter to $108 billion, with balanced growth across hyperscale and AI cloud, industrial, and enterprise subsegments.
NVIDIA’s management sees a backlog of $2.5 trillion in the cloud industry in 2026 Q3 (FY2027 Q3), up from $2 trillion in 2026 Q2
Cloud industry backlog now sits at $2.5 trillion, up from $2 trillion just three months ago, underscoring the significant multi-year opportunity ahead of our partners and of NVIDIA.
NVIDIA’s management is seeing rapid revenue growth at the frontier labs; management expects the frontier labs to account for 15% of NVIDIA’s data center revenue in 2026 (FY2027)
As frontier models become more capable, adoption and revenue are scaling rapidly, with Anthropic’s annualized revenue reaching $81 billion at the end of October. We expect the major frontier labs to contribute approximately 15% of our data center revenue this year.
Weekly token volume on OpenRouter is up 35x year-on-year, and 3.5x sequentially, in 2026 Q3 (FY2027 Q3); NVIDIA is acquiring Hugging Face, an open model community serving more than 18 million developers and 200,000 companies; NVIDIA’s Nemotron is the most widely used open model on OpenRouter; management is seeing enterprises use post-trained open models; management is seeing training partners help enterprises turn open models into their own; management is seeing startups use post-trained open models in vertical domains
Open models are accelerating AI adoption, and weekly token volume on OpenRouter has surged 35x year-over-year and 3.5x in just this past three months…
…In September, we entered into a definitive agreement to acquire Hugging Face, the home of open model community, serving more than 18 million developers and 200,000 companies…
…Nemotron is now the most widely used U.S. open model on OpenRouter and is where developers start to build their own intelligence. That foundation is empowering the ecosystem.
This past quarter, leading enterprise software and security companies, including Microsoft, Salesforce, Palantir, CrowdStrike, Cadence, and Synopsys announced products that use post-trained open models tailored to domain-specific use cases. Training partners such as Scale, Mercor, Prime Intellect, and Applied Compute are helping enterprises turn open models into their own using their data, tools, and workflows. And startups across specialized domains, including Lila Sciences, Harvey, Therma, and Abridge, are using post-trained open models to deliver results in science, law, and healthcare.
Agentic workloads now account for 80% of tokens in production; management has created the NVIDIA Open Agent Safety Platform to enhance the security of agentic workloads, and it is already being adopted by large enterprises; Meta Platforms launched its Muse agent product with NVIDIA’s infrastructure
Agentic workloads now account for approximately 80% of tokens…
…As agents run longer and act with greater autonomy, organizations need stronger controls to ensure safety and security. The NVIDIA Open Agent Safety Platform, which we announced alongside more than 100 partners, enable these controls. OpenShell contains agents by limiting their access to files and tools, while NVIDIA Sentry, running on BlueField-4, monitors the behavior in silicon and stops agents for overstepping limits in milliseconds. OpenShell is already gaining broad ecosystem support from partners across AI models, enterprise software, cybersecurity, cloud infrastructure, and robotics. Companies working with NVIDIA include Anthropic, AWS, CoreWeave, CrowdStrike, Dell, Google Cloud, HPE, IREN, JPMorgan Chase, Lenovo, Microsoft, Nebius, OpenAI, and SpaceXSI.
Advances in agent harnesses are accelerating the adoption of agentic AI. Meta’s Muse agent, launched on NVIDIA’s infrastructure, surpassed 10 million setups in the United States and Canada within the first six weeks.
NVIDIA’s management saw token prices fall 60% since May 2026, and this has accelerated demand, leading to higher GPU prices across generations
Token prices have fallen nearly 60% since May, accelerating token usage and demand for compute. As a result, GPU rental prices are rising across generations. Hopper GPUs have recontracted at up to twice their original price under the multi-year agreements, while the B300 rates have risen more than 60% since their introduction. NVIDIA Compute remains fully utilized across every hyperscaler and neocloud we serve, underscoring the productivity, durability, and fungibility of our infrastructure.
NVIDIA’s management sees agentic AI driving strong demand for CPUs; CPU supply cannot keep up with demand; NVIDIA’s Vera CPU is being adopted by many important AI companies, and management sees very strong demand
Agentic AI is driving extraordinary demand for CPUs, far outstripping supply. NVIDIA Vera CPU meets the growing demand, delivering leading performance at data center processing and scientific simulation workloads. CoreWeave, Meta, Nscale, and OpenAI have now adopted Vera, joining AWS, OCI, and SpaceXSI, the leading partners we announced last quarter. We continue to see very strong demand for Vera and expect total CPU revenue to more than double in fiscal year 2028.
NVIDIA’s management thinks neoclouds are increasingly important players in the global AI infrastructure ecosystem; NVIDIA has been supportive of neoclouds, and management intends to be disciplined and selective in its support; the area of support management wants to provide for neoclouds is in land, power, and shell; NVIDIA’s current guarantees that are related to neoclouds will only kick in when the neoclouds’ data centers are ready for production
Neoclouds remain an increasingly important part of the global AI infrastructure ecosystem, providing additional customers with access to NVIDIA Compute and expanding the reach of our platform. We have supported the development of this ecosystem, and we will continue to be disciplined and selective in our support…
…[Question] ACIE is a very important growth vector for NVIDIA. I am hoping you could update us on your appetite for future vendor financing, and if there is any change here, what are the implications for growth for this important segment,
[Answer] How do we think about in terms of the future and potential financing that will be necessary? The most important part of the financing that we would probably review, but have to review in detail to make sure it is the right thing, is focusing on land, power, and shell. The availability of land, power, and shell is such a critical part of the whole structure in terms of building out those data centers, so they may need the assistance. That’s a great asset for anyone to hold, so we will look at that as probably one of our top priorities… We can look at them in terms of the assistance of needing in terms of prepaying, which can help in terms of their build of that data center. Right now, I think we would look at this and taking a very strong look in terms of the types of companies that we would want to help in terms of the financing, as a lot of things have changed over this last quarter, and we know that that’s a priority of ours…
…[Question] In the CFO commentary, it looks to me a new sentence, a new paragraph on the neoclouds. It says, “Your commitments take effect only if the neoclouds secure financing to purchase the data center infrastructure products and related compute capacity deployed, maximum commitment,” blah, blah. It says, “Certain commitments may be reduced or eliminated,” if they don’t do this. I guess what I’m asking is, it’s a new sentence in the CFO disclosure. Is this something new in how you’re doing business with the neoclouds, or is this something that you’ve been doing all along and you’re just telling us about it now? Because it sounds like you guys are not the ones that are actually looking to provide the financing. It reads like they have to secure their own financing in order to get the backstops and commitments from you.
[Answer] When we are looking at many of our different types of guarantees, much of the land, power, and shell that has to be built and put together, there is construction, there is the raising of capital that they may need through equity and/or debt in order to build out any of the buildings, get them ready for use, but also the ability to procure the compute, install the compute, and get ready. What we are indicating in our statement there is that is when our guarantee starts. If all of that begins and unfortunately they are delayed or they have a challenge in terms of building that, our guarantee doesn’t exist anymore. Our guarantee for most of these are really about when they are ready to go, particularly when we think about our one in terms of Portsmouth, in terms of what we could do. These are important things to understand because that shows what we have done in terms of really thoughts, in terms of very, very strong businesses that we felt could actually maintain creating the debt and equity that they needed to get that done and create the data center such that we felt very confident in terms of that guarantee. Is it new? No. Did we change anything? No.
NVIDIA’s management recently expanded its Isaac robotics platform; NVIDIA is working with Mercedes-Benz, Lucid, and Wasabi, to develop autonomous vehicles
At GTC Berlin, we expanded our NVIDIA Isaac robotics platform with a new humanoid reference design built on Isaac GR00T and Jetson Thor. In automotive, NVIDIA Hyperion and Alpamayo are enabling Mercedes-Benz and Bolt with Lucid to build robotaxi fleets, and Waabi to develop level four autonomous trucks, and Volvo Cars to advance safe automated driving.
NVIDIA earned minimal revenue from China in 2026 Q3 (FY2027 Q3); management does not include China in its outlook
Q3 sales of licensed Hopper data center products to China was $2.4 billion. Our guidance does not have any licensed Hopper data center products included, and we are no longer shipping licensed Hopper data center products to China.
NVIDIA’s management has signed on more purchase commitments; NVIDIA’s purchase commitments is $390 billion in 2026 Q3 (FY2027 Q3), up from $279 billion in 2026 Q2; NVIDIA’s data center lease and cloud service commitments is more than $100 billion in 2026 Q3 (FY2027 Q3), up from $45 billion in 2026 Q2; NVIDIA has $105 billion in credit support commitments for land, power, and shell capacity for an AI data center that will be leased to OpenAI for 20 years; Softbank Energy’s plans for the OpenAI data center are on track
With demand continuing to outpace supply, we are working closely with our partners to accelerate capacity expansion across critical components, including wafers, substrates, and advanced packaging. Today, we are announcing new agreements with Kinsus and Samsung Electro-Mechanics in substrates, and SPIL in advanced packaging, further strengthening our global supply chain capacity and resiliency. These build on partnerships we announced last month with Amkor, IBIDEN, KYEC, and Unimicron. In Q3, our purchase commitments increased to $390 billion, reflecting our confidence in the long-term sustainability of demand for NVIDIA infrastructure. We are securing data center and cloud capacity required to support our R&D efforts across hardware and software, and increasingly in open models.
Together, our data center lease and cloud service commitments increased to a little over $100 billion, driven primarily by our long-term leases. Commitments associated with our neocloud agreements totaled $32 billion at the end of Q3…
…Total guarantees were $118 billion at the end of Q3, including the $105 billion in credit support for land, power, and shell capacity for the PORTS-Pike campus we discussed last quarter. The site will exclusively host NVIDIA infrastructure under a 20-year lease to OpenAI. SoftBank Energy plans are on track, with capacity expected to come online in phases beginning in calendar 2028.
NVIDIA’s management continues to work with capital partners to provide financing for the build-out of AI infrastructure
We continue to work closely with our infrastructure capital partners to stand up the financing platforms. These partners are instrumental in financing the build-out of AI infrastructure. As we advance this work, we will remain thoughtful in how we participate and disciplined in how we commit capital. We are proud of the role that we have played so far in supporting the ecosystem. As the market matures and additional sources of capital become available, we are rigorously evaluating future investments, commitments, and guarantees.
NVIDIA’s networking revenue exceeded $21 billion in 2026 Q3 (FY2027 Q3); NVIDIA’s co-packaged optics (CPO) roadmap is on track, with CPO scale-out in volume production; Lambda is the first company to deploy NVIDIA Photonics CPO switches by connecting 10,000 GB300 GPUs; NVIDIA Photonics switches has 5x better power efficiency, 15x higher reliability, and 10% lower latency, compared to pluggable optics
Our networking franchise continues to scale, with Q3 revenue exceeding $21 billion. Our co-packaged optics roadmap remains on track, with CPO scale-out in volume production. At Supercomputing 26, Lambda announced the first production deployment of NVIDIA Photonics CPO switches, connecting more than 10,000 GB300 GPUs. Compared with pluggable optics, NVIDIA Photonics switches delivered 5x better power efficiency, 15x higher reliability, and 10% lower latency.
NVIDIA recently introduced the DSX POD, a modular AI factory reference architecture; the DSX POD cuts deployment times of GPU systems from months to weeks, potentially generating materially more revenue for data center operators; NVIDIA is working with more than a dozen partners for the DSX POD, and plans to ship the product in early-2027
At GTC Berlin, we introduced DSX POD, a full-stack modular AI factory reference architecture. Just as Grace Blackwell NVL72 moved rack assembly from the data center into the supply chain, DSX POD extends that approach to the entire cluster. Pods arrive ready to connect, cutting deployment times from months to weeks, accelerating time to first token. With annual revenue generating from one 1GW AI factory approaching $50 billion, reaching first token just one quarter earlier could generate more than $10 billion in additional revenue for the operator. We are working with more than a dozen partners, including Dell, Foxconn, Schneider Electric, and Vertiv. We plan to ship our first pods to customers in the U.S. and Central Europe early next year.
NVIDIA’s management thinks the company’s gross margin will increase in the future, despite the frontier labs leaning into their custom AI chips
[Question] I wanted to see if you can speak to your ability to maintain pricing power and margins over the longer term. I ask because if you look at the frontier model guys, they’re leaning in even more aggressively on ASICs, and they’re talking about compute efficiency as new KPIs.
[Answer] In our last earnings, we did discuss where we saw our gross margins, not only for the last two quarters of FY 2027, but we also gave an outlook in terms of FY 2028. If you recall at the time, there was a significant change in terms of the cost of supply throughout the quarter that we were in, as well as what we were going to move in terms of FY 2028. So it was important for us to rework in what we thought would be important. Yes, we thought about in terms of our prices and pieces of that, and a lot of the costs were some of the things that we absorbed in those numbers that we included. What we provided here in Q4 is our bottom.
NVIDIA’s management thinks the blue-sweep of the US midterm elections will not affect the demand picture for AI infrastructure in 2028
[Question] I’m sure you’re aware, the big fears that there could be a slowdown after your great calendar 2027. Is there something that you’re seeing increased confidence in this? Have you spoken to anyone in the administration, or not administration, but in government since the blue wave, and anything that gives you confidence beyond 2027 that you might be able to tell us about?
[Answer] Where we stand right now, there is so much demand with such the largest companies continuing to work on so many agentic solutions, new models that are also available, and that’s been our focus to assist. I’d say even in the last week, the continued ask, how are we going to receive more and more supply to help them with the demand that they have in front of them? So they’re all getting ready for that. I know there’s been a lot of focus in terms of the data centers within the U.S. as well. Are they well-positioned to take care of the communities that are around that? think that’s very important that all companies are working together so that communities understand both the jobs that are going to be newly available, but giving them the confidence regarding how they will be perceived and working within those areas as well. I think that is an area to focus on, but I think there is a lot of data centers already in use or ready to go that might be that case. We’re always going to be engaged with the government, making sure that we are all aligned in terms of what we see in terms of the future. But I think the demand and what we have in terms of agentic will continue not only in the 2027 perspective, but moving to the 2028 as well, as we’ll continue to carry over the amount of demand that’s in front of us.
Sea Ltd (NYSE: SE)
Shopee is an early connector partner of Meta’s Muse AI personal agent that recently launched in Singapore; management thinks Muse could become a solid additional channel for buyer acquisition and incremental GMV growth over time; Shopee is also a partner of ChatGPT in agentic commerce; there’s currently no difference in the economics that Shopee enjoys when any agentic-driven transaction happens; management sees the partnerships with the AI agents as similar to the partnerships Shopee has with affiliate partners
Shopee is also one of the early connector partners for Meta’s Muse AI personal assistant that has just launched in Singapore. Users will be able to research and compare products on Shopee through Muse, helping them make purchase decisions tailored to their preferences. The partnership is still at an early stage, but we see potential for it to become a solid additional channel for buyer acquisition and incremental GMV growth over time…
…On the agentic AI, we actually partner with a range of partners, starting from ChatGPT, and now, as you mentioned, the Muse. At this point in time, all the checkout and seller relationships and buyer relationships happen within the shopping platform. We didn’t see a meaningful impact to our economics for the near future. At the same time, I would like to point out that we work with many platforms as affiliate partners as well. For example, we’ve been working very closely with both YouTube and Meta for Facebook and Instagram as affiliate partners, which works similar to the fashion we work with platform like Muse. Affiliate partner, we’ve been generating a sizable new user to our platform, generating a sizable traffic to our platform, and it’s been part of our economics in the past years. We don’t see this will dramatically change the picture when we work with Muse or ChatGPT.
Sea’s management announced a new business, Sea Cloud, recently; Sea Cloud provides infrastructure, model and agentic applications for AI, 3 areas that Sea has built expertise in the past 2 years; management thinks the real prize in AI is to take advantage of the long-term decline in token prices and help organisations drive much broader broader AI adoption; management sees lasting value in helping organisations (1) run AI efficiently, and (2) put AI to work, including shifting to the most cost-efficient models as and when they emerge; management sees Sea Cloud as an opportunity to turn Sea’s AI capabilities, which are currently a cost for Sea, into a profit center; management’s current focus with Sea Cloud is on software and customer relationships, and not on owning large amounts of compute capacity; if Sea Cloud’s customers require compute capacity, management will build the capacity only if there are clear customer commitments and attractive returns; management currently does not expect Sea Cloud to have a significant impact on Sea’s financials; management wants to be very frugal with capex for Sea Cloud in its early days; management is aware of the many unknowns for Sea Cloud; Sea Cloud has yet to secure any enterprise contracts; management will be evaluating Sea Cloud on a deal-by-deal basis
In addition to our three existing business segments, we have always been exploring and experimenting with new businesses. The most recent of these is Sea Cloud, as we announced recently. As AI has taken off, we have built up meaningful capabilities over the past two years across the infrastructure, model and agentic application layers, and our own businesses have benefited significantly from them. Sea Cloud looks to bring these capabilities to other organisations across all three layers.
Today, there is a lot of focus on compute prices and cyclical margins, and having enough compute certainly matters. But we believe the ultimate prize lies elsewhere. The cost of tokens is falling rapidly and structurally, and as it does, AI becomes affordable for far more organisations and far more use cases. The real opportunity is to build a significant business around the long-term decline in token prices, helping organisations drive much broader AI adoption.
We see much of the lasting value in the layers above bare-metal capacity. One is running AI efficiently: getting more out of every GPU through better utilisation, faster deployment and leaner operations, which we have mastered for our own businesses. The other, and the more important over time, is the agents, software and orchestration that help organisations put AI to work: running models reliably and securely, adapting them to their own data and languages, keeping that data close to home, and moving to newer and more cost-efficient models, including open models, as they emerge, without having to rebuild each time. Many organizations want this, but few have the specialist teams to build it themselves. That is where we intend to focus, and it is not theoretical for us. It is what we have had to do for our own businesses, running AI at scale in areas such as fraud detection, customer service and content generation, under real commercial pressure and, in Monee’s case, in highly regulated environments. We see this as an opportunity to turn our AI capabilities, today a cost centre for Sea, into a profit centre…
…We are setting up Sea Cloud as a frugal start-up, so that it can stay nimble and respond quickly to market demand. Our focus is the software and customer relationships, not owning large amounts of physical capacity. Where customers need dedicated computing capacity, we will only pursue opportunities with clear customer commitments and attractive returns. We will mainly look to partners and customer prepayments to fund the physical infrastructure. As a result, we do not expect Sea Cloud to have a significant impact on our financials for now. It is still early and we will update investors on any meaningful developments…
…At the same time, because of this is very new and it is a very dynamic situation and opportunity, we want to make sure we will be very, very frugal in terms of the CapEx, in terms of the cash flow spending…
…There is a lot of things we do not know yet. Again, this is very early. This is a kind of a startup. There is a lot of things we need to figure out. For example, what exactly the segment we should focus on, what exactly type of customers we should focus on, and what is the best business model we should use. We think we will get a better picture when we really get our own feet into the game and we start to get first-hand experience…
…I do not think at this moment we have a very, very clear as overall as a business what will be the return on investment will look like. We believe this will be very, very good. But again, we also want to stay humble. There is a lot of things we do not know what we do not know…
…for Sea Cloud, for your question, as you know, this is just at the very, very beginning of the business. At this moment, we haven’t secured any enterprise contracts…
…I think it will be very much deal-by-deal driven, at least at the beginning stage, and with very, very specific target in terms of the customers, in terms of the projection of the revenue, and what is the timeline for that, and what is the potential investment we put in. I think we are going to have a guardrail. We want to, as I mentioned in the remark, we’re only pursue opportunity with a clear customer commitment and a clear return profile. We’re going to make the case-by-case judgment if that return is good enough for us or not. We don’t have to pursue the opportunities if we don’t like the return profile.
Tencent (OTC: TCEHY)
HY4 (Hunyuan 4) is now a state-of-the-art model in China; HY4 Flash lowers inference cost for production workloads; Tencent scaled up pre-training and post-training for HY4 and added vision and computer use capabilities, leading to HY4 performing much better in instruction following, long context understanding, long horizon tasks, and deep research compared to Tencent’s earlier models; HY4 Flash is priced at 5% of HY4; HY4 Flash’s usage has exceeded management’s expectations; HY4 Flash has the highest daily token consumption on OpenRouter of any model launched in 2026; HY models are deployed in Tencent’s key products, which provides valuable signals for the company to improve the models; management recently unified multi-modal model development under the Hunyuan team; Tencent’s models are industry-leading for 3D generation; Tencent’s model team was rebuilt from scratch, and pushed HY4 to state-of-the-art within a year; a good chunk of HY4 usage is outside China, but Tencent can serve a part of that usage from within China; management wants to serve HY4 around the world
Hy4 has achieved China’s state-of-the-art AI model status, and Hy4 flash brings down inferences cost to deliver cost performance for real-world use cases…
…The model ranked first among Chinese models based on Agent Arena leaderboard at launch, achieving China’s state-of-the-art level…
…Our team substantially scaled up pre-training and post-training for Hy4 and added vision and computer use capabilities. As a result, Hy4 delivers significant gains in instruction following, long context understanding, long horizon tasks, and deep research compared to our earlier models…
…Hy4 flash is priced at 1/20th of Hy4. Its usage way exceeded our expectation and achieved the highest daily token consumption volume on OpenRouter of any model launched this year…
…By deploying Hy models in key products, including WorkBuddy, we capture high-value user feedback, enabling continuous model-product co-design and a virtuous iteration cycle, especially for real-world use cases. We have also recently unified multimodality development under Hunyuan team. We lead the industry in terms of 3D generation, and we are improving image generation and also developing video generation capabilities…
…I think the fact that our model team was actually completely rebuilt from scratch, and within a year, they were able to push our model to top-tier position in China…
…We also see pretty good chunk of the usage of Hy4 is actually outside of China. In order to serve that, we could serve a part of that from the inferencing capacity in China. Over time, we actually want Hunyuan 4 to be promoted around the world, and we may find partners to do that. There are inferencing partners around the world who can actually do that.
WorkBuddy has become the largest AI application in China by token consumption; WorkBuddy now has 10 million DAUs (daily active users) and they consume 100 trillion tokens per day; WorkBuddy is a clear leader in the productivity AI domain; management thinks adding AI capabilities to productivity software will significantly increase spending on such software for 2 reasons, (1) avoiding piracy, and (2) delivering many valuable use cases in a single package; management thinks WorkBuddy’s subscription plus usage-based monetisation model can unlock a much larger revenue pool, similar to how China’s online game industry evolved; WorkBuddy’s annualised gross revenue in the last 7 days exceeded US$500 million; management has seen enterprises scale their spending on WorkBuddy significantly when FDEs (forward-deployed engineers) help demonstrate value; 60% of the top 100 companies in China have become WorkBuddy customers since the enterprise version of WorkBuddy was released in June 2026; management is confident that WorkBuddy’s unit economics will improve over time; WorkBuddy’s revenue was up 5x sequentially in 2026 Q3; WorkBuddy is the biggest customer today of almost all of China’s open-weight models; management wants WorkBuddy to grow outside of China
We believe WorkBuddy has grown to become the largest AI application in China, measured by total daily large language model token consumption by users…
…WorkBuddy has grown very rapidly to over 10 million daily active users, consuming over 100 trillion tokens each day…
…WorkBuddy is also by far the leader within productivity AI domain, and according to external surveys, internal data has widened its lead versus peers in recent months.
While we have historically been cautious about how quickly the enterprise software market would develop in China, we now believe that adding AI capabilities will allow productivity software spending to rapidly and substantially increase, with WorkBuddy as a key driver and beneficiary. The rationale is as follows. Firstly, AI services avoid software piracy problem. Secondly, WorkBuddy covers a large number of valuable use cases within enterprises, such as document preparation, data analysis, project management, research, and software development, which makes the whole package worth paying for. The subscription plus token usage-based monetization model means that customers can unlock a paid experience at a low flat entry price, but then spend more as they derive more value. WorkBuddy’s annualized gross revenue over the last seven days has exceeded US$500 million, which is more than 2x of Tencent Meeting, which was launched seven years ago.
In some ways, these dynamics mirror those of China’s online game industry evolution, where cloud connectivity also curbed piracy, and an item-based monetization model unlocked a much larger revenue pool when spending can scale up along the demand curve. We have already seen enterprises significantly scale their spending up when forward deployment engineers demonstrate how usage can add value. We are working with partner companies to provide FDE support to WorkBuddy customers. Since we released the enterprise version in June, 60% of the top 100 companies in China have become WorkBuddy’s customer. This strong early adoption is very encouraging, and we believe this is just the tip of the iceberg.
While we are currently investing to grow the WorkBuddy adoption, including via providing temporary free services to certain categories of users, such as school teachers, we are confident that we will progressively improve the unit economics for WorkBuddy over time by increasing the paying user ratio, facilitating higher token spend per paying user, and managing down token costs via inference optimization, model customization, and compute efficiency improvement. By integrating WorkBuddy with Tencent and external ecosystems, we believe we can enhance WorkBuddy’s user retention, lifetime value, and overall economics…
…For WorkBuddy, revenue for the third quarter was over 5x the previous quarter, driven by increasing paying users and strong initial enterprise adoption…
…When we look at WorkBuddy as a product, then WorkBuddy works with very efficient models, first of all, but it also works with a whole range of different models, and it gets a very diversified supply of tokens, some of them provided by us, some of them actually provided by the model companies, the various model companies that WorkBuddy hosts. As a matter of fact, WorkBuddy has become the biggest customer of almost all of the open-weight models in China.
Tencent’s management thinks its personal assistant product Xiaowei, which is embedded in Weixin, will help diffuse AI into society; Xiaowei is under testing now but already has 30 million DAUs (daily active users); management has expanded the number of touchpoints Xiaowei has within Weixin to more than 30; management is working to enable Xiaowei to facilitate transactions within Weixin; inference costs to run Xiaowei are very low; management thinks Xiaowei will cost less to run, even with 1.4 billion DAUs, than Yuanbao was incurring; management thinks Xiaowei will only become more capable over time as compute efficiencies improve; management thinks Xiaowei’s costs can be further reduced over time when some inference workloads are run on devices; Xiaowei is able to integrate with merchants at minimum cost; management expects Xiaowei to be a monetisation multiplier on Weixin’s existing revenue streams, with advertising and transaction-based revenues being the immediate opportunities; management also thinks Xiaowei can add new revenue streams for Weixin over time; the number of interactions per user among Xiaowei’s DAUs has been increasing; Xiaowei is not as capable as Meta Platforms’ Muse, but this is by design because management wants to keep the cost of serving Xiaowei low; management expects Xiaowei to be as capable as Muse eventually, while serving the entire Weixin userbase; management expects Xiaowei to be a hugely profitable product
Xiaowei, the personal assistant embedded in Weixin, will facilitate AI diffusion breadth and make AI accessible to every single internet user in China. Since testing started in June, we have consistently expanded the user testing pool for Xiaowei, and now has achieved 30 million DAU. We have added more Xiaowei touchpoints into Moments, Official Accounts, and Video Accounts, unlocking greater utility across a broader range of scenarios and driving higher interaction per DAU. We are in the process of onboarding major Mini Programs, merchants, and integrating with Mini Shops in order to facilitate transactions…
…In terms of Xiaowei, we have been gradually rolling out to more and more users, and we have also been expanding the number of touchpoints within Weixin to now more than 30 touchpoints, including Moments, Official Accounts, Video Accounts…
…Because Xiaowei is powered by the purpose-built VLM model, customized for cost efficiency and optimized for the Weixin ecosystem, inference costs to run Xiaowei are very low and will continually trend down. In fact, even when Xiaowei is available to the entire 1.4 billion Weixin users, we expect the quarterly cost of Xiaowei will remain lower than the quarterly cost that Yuanbao was incurring earlier this year. Xiaowei’s capabilities will be significantly enhanced as compute efficiency multiplies every year. Over time, some inference workloads can be processed on user devices, and that would further reduce costs. By leveraging Weixin’s pre-existing network of merchants across payment and Mini Program services, Xiaowei is able to achieve powerful integration with merchants and service providers at minimum cost, and to deliver synergies rather than competition with them, fostering win-win relationships.
On the monetization front, we expect Xiaowei to act as a multiplier that we apply to Weixin’s already substantial revenue streams, compounding them further and faster. Advertising and transaction-based revenues are immediate opportunities, and we can add new revenue streams such as agent-to-agent facilitation services revenue from merchants and service providers over time…
…Our DAU in terms of interaction with Xiaowei is now 30 million, and the number of interactions per day has been increasing as well…
…A lot of people, if you compare it against Muse, would say, oh, it’s not as powerful as a Muse, which is the case by design. Because we actually want Xiaowei to first come out not as all-capable agents available to a small number of people who are willing to pay for it. We actually want Xiaowei to be available to all Weixin users, and over time, it will become more and more powerful because of the connection with the different touch points, because of the connection with the ecosystem, including the Mini Programs. We actually need to take time to have more Mini Programs to have CLI interaction with our integration with our Xiaowei. A lot of the capabilities would actually sort of be built over time. At the same time, because of the cost concern. Because if we want to make it available to all Weixin users, then there is a cost question. That’s why we use a small model, and we’re confident that even if we roll out to all users, it will not be such a big cost…
…At some point in time, it will be as powerful as Muse. But we can actually cover all the users within Weixin…
…If you look at the current revenue stream of Weixin, which is, if you combine everything, in the hundreds of billions of RMB a year, right? Then adding 10% or 20% to that would be a very significant number, which will make Xiaowei hugely profitable for us.
Developers can now use WorkBuddy to automate the development and publishing of Mini Programs in Weixin; management has introduced AI harness for creators to generate more content from their Official Accounts articles
Developers can now use WorkBuddy to automate their development and publishing of Mini Programs, lowering entry barriers and shortening development cycles. For Video Accounts, we boosted engagement among younger users via deeper content collaboration with our internal game, music, and TV services. We introduced an AI harness for creators to convert Official Accounts articles into videos, automating storyboarding, scripting, and media generation.
Tencent’s management has introduced AI-powered advertising solutions for content publishers; the AI solutions include automated insertion of ads for mini-drama providers that come with optimal length at optimal times; the AI solutions now power 80% of ad monetisation for mini drama publishers
We introduced AI-powered advertising solutions for content publishers. For example, we enabled mini drama providers to auto-embed advertisements of optimal length at optimal times, minimizing disruption to user viewing while enhancing monetization efficiency. These AI solutions now power approximately 80% of ad monetization for mini drama publishers. Video Accounts ad impressions grew over 50% year-on-year, driven by video views and ad loads. We serve advertisements based on our content ranking system now, which increases our ad recommendation relevance and monetization.
Tencent’s management is starting to accelerate the addition of compute capacity and have made some of the capacity available to external customers of Tencent Cloud; within the Fintech and Business Services segment, cloud services revenue grew 35% year-on-year in 2026 Q3, up from 21% growth in 2026 Q2, and teens-level growth in 2025; WorkBuddy’s revenue was up 5x sequentially in 2026 Q3; Model as a Service revenue was up more than 3x sequentially in 2026 Q3; GPU rental services revenue was up more than 30% sequentially in 2026 Q3; management does not expect tougher US restrictions on AI chips to impact the company’s cloud business; Tencent’s cloud services is primarily running inference, and is doing it effectively on a limited number of domestic chips; management is unable to offer high-end GPUs anymore via the cloud services business
We are at the early stages of adding compute capacity at an accelerated pace and making a portion of that capacity available to external paying customers through Tencent Cloud. As a result, our cloud services business is now starting to deliver financial benefits from our AI initiatives, initially from WorkBuddy, Model as a Service, and GPU rentals. During the third quarter, cloud services revenue growth accelerated 35% year-on-year from 21% year-on-year in the second quarter, and teens level percentage year-on-year growth throughout 2025.
For WorkBuddy, revenue for the third quarter was over 5x the previous quarter, driven by increasing paying users and strong initial enterprise adoption. For Model as a Service, revenue more than tripled quarter-on-quarter, with enhanced inference efficiency and strengthened operational controls and security guardrails for enterprises so that they have oversight of their AI deployment. Our GPU rental services revenue grew more than 30% quarter-on-quarter, driven by enterprise demand for inference and improved scheduling efficiency for our compute capacity…
…[Question] How do increased U.S. restrictions impact both existing and future deals with compute, and does that impact the acceleration we saw in cloud and elsewhere in the business going forward?
[Answer] As far as how the restrictions impact our cloud business, we actually don’t see an impact because our cloud business today is primarily running inference, and in particular, running inference for highly efficient models such as Hy4 flash and other Flash models. We’re now at a stage where that inference can be run very effectively on a surprisingly limited number of domestic chips. I think in the last 12 months, there’s been some transformations in terms of the inference productivity, and consequently, we’re able to drive a far higher volume of tokens, specifically on domestic chips, than we would have imagined a year ago…
…In terms of cloud, if it is non-GPU related, we can actually build capacities ourselves outside of China. I think there is still a lot of such opportunities because our cloud service is actually very cost-competitive. When it comes to GPU, then unfortunately, we cannot actually offer high-end GPU anymore. We may be able to work with some partners in different places.
Tencent’s operating capital expenditure in 2026 Q3 was up 257% year-on-year, and up 23% sequentially (was CNY 51.8 billion in 2026 Q2) because of accelerated investments in AI infrastructure; free cash flow was deeply negative in 2026 Q3, and down year-on-year and sequentially; Tencent’s operating capital expenditure in 2026 Q3 included significant prepayments for compute procurement; free cash flow would be positive if the prepayments were excluded; the prepayments in 2026 Q2 and 2026 Q3 were abnormally high, and management expects prepayment outflows to decline going forward; management thinks cash capex is at peak levels now, and will reduce from 2027 onwards; management expects the prices of AI infrastructure to come down quite significantly in 2028; the existence of HY4, HY4 Flash, and WorkBuddy, is proof that Tencent’s capital expenditure for AI is generating very valuable products; management is uncertain when Tencent’s AI products will generate a profit, but they are confident it will, eventually
Operating CapEx payments were CNY 63.6 billion, up 257% year-on-year, as we accelerated investment in AI infrastructure to support enhancement to our Hy model family, WorkBuddy inference needs, and development of AI capabilities across our products and services, as well as to meet external demand for our cloud services. Non-operating CapEx payments were CNY 1.5 billion. Free cash flow was CNY -69.1 billion, mainly due to large compute procurement. Excluding prepayments from compute procurement, our free cash flow would have been CNY +16.9 billion…
…Regarding prepayments, the second and third quarter were abnormally high levels, and we have now entered into the long-term contracts that we wanted to enter into, so our prepayment outflows will decline. In fact, there may be some quarters where prepayments reverse due to the cancellation of contracts that we previously prepaid, but we’re now exiting as a result of recent changes.
In terms of the CapEx separate from the prepayments, the current run rate, we’d look at it on a cash CapEx basis, which I think was CNY 65 billion, rather than the CNY 81 billion accrual CapEx, because if you’re looking at prepayments and CapEx together, then you should look at prepayments and cash CapEx. Otherwise, you’re double counting looking at prepayments and accrual CapEx. Anyway, looking at the cash CapEx level, we think that that is at peak levels, and cash CapEx may bounce up and down around this peak level through the rest of 2026 and 2027 on a quarterly basis, and then we expect it to reduce beyond there…
…The pricing that we actually see in 2026, 2027 is actually elevated. As we go into 2028, even if you need to buy the same thing, you should expect the price should come down quite significantly…
…Before our Hy3 and Hy4 got launched, it’s uncertain whether our training investments is generating the valuable models. But now with Hy4 and Hy4 flash coming out and we see the explosive endorsement of users by using up a lot of tokens, I think we put a check on it and say, oh, the investment in training is now generating very valuable products…
…The next product that we actually launched was WorkBuddy, and WorkBuddy has become a phenomenal success. So now we can actually put a check and say all the investment that we put around AI products is generating positive value…
…Over time, we know these products will generate, first of all, usage, then revenue, and then, over time, profit. Exactly when would that happen? No one knows. We are really operating in a generational technology revolution. I think it’s very difficult to build a business like, you build a building and try to get rental return from it. But when we look at the investment that we made and we’re building very valuable results, then we feel confident that over time it will generate revenue and it will generate profit eventually.
Tencent’s management is seeing the company’s compute efficiency multiply quickly
We now, after working on the training and inference of our models, we see that compute efficiency is actually multiplying quite quickly year over year. That means we’re going to be able to get higher training efficiency as well as a much larger number of tokens as we go by the year with the same facility.
Tencent’s management thinks the latest US restrictions on AI chips sets everyone back in China, especially when training AI models; management expects the restrictions to increase the time Tencent needs to train its models in the short-term, but in the long-term, there’s little impact because domestic chip-supply would be coming onboard; management thinks all Chinese companies building AI will have to rely on domestic chips eventually; management expects the latest US restrictions to speed up the development of China’s domestic chip supply chain
When we look at the impact on our AI business in relation to the latest tighter restrictions, I would say it’s unfortunate, but a few points. Number one, it does set everyone back. It sets everyone back in China. So everyone is actually having less compute, advanced compute to work with, no matter when you actually started leasing or building capacities outside of China. Number two, when we look at training, then it’s going to be impacted. But our training efficiency with Hy4 is actually very high. When we look at the next version, it would take probably longer time for us to come up with the next version in the short run. But over time, the domestic supply chain would actually come on board…
…Over the longer run, we believe all Chinese players would actually have to rely on domestic supply chains and domestic chips, which we actually see the industry is actually progressing very fast. With the latest restriction, we expect it to actually speed up even further, because all the development effort would actually have to be focused on speeding up the domestic supply chain.
Tencent’s management thinks AI can be applied to the company’s businesses, but a major reorganisation of the company is needed; management thinks AI can reduce the people needed for running existing businesses, allow small teams to build new businesses, and thus allow the freed-up headcount to build new growth drivers
AI really can be applied to various businesses to improve the efficiency of how businesses are run. But in order for you to be able to do that, you actually need to have a pretty significant reorg of the organization, and also you need to re-engineer a lot of the work processes. If we can actually do that, right, then existing businesses can actually use fewer people, right? But then at the same time, what we see is AI actually allows small teams to leverage AI to explore new ideas and build new businesses and build new growth drivers, right? If we can actually use fewer people for running an existing business and actually can redeploy these people to build new growth drivers for the existing business as well as for building new businesses, right, then I think that will be the ideal, right?
Tencent’s management thinks the game environment broadly has seen innovation stall, but management is hopeful that AI will drive re-acceleration; management sees AI as both a threat and an opportunity for Tencent’s games business, but as of today, AI is unable to take over the production process of the type of high-quality games that power Tencent’s business; management thinks AI can expedite the evolution of many of Tencent’s evergreen games into evergreen game platforms
The game environment globally is one where the rate of innovation has slowed somewhat, and now with AI, there’s opportunities to re-accelerate innovation, and that currently is more around the production of in-game content and the pre-production for new games rather than full production, post-production for new games…
…AI can be a threat or an opportunity for the game business, for the game industry globally. We have seen that in some categories of games, such as hyper-casual, single-player mobile games, that there have been rashes of AI-created copycat games that have arisen very quickly as a result of people running decompilers on existing games. Recently, the big model laboratories have taken steps to prevent that decompilation or piracy process, which seems to have curtailed the AI copycatting. But we are very early in the evolution. I would say that the work involved, the expertise required to not only create, but also operate and sustain and grow over time a multiplayer evergreen game that is multi-platform, multimodal, is orders of magnitude greater than the work required to create a single-player hyper-casual mobile game. Obviously, those multiplayer gameplay-based, multi-platform, multimodal evergreen games, where we make most of our game revenue. AI can accelerate aspects of that process, as I mentioned in an earlier question, such as the pre-production, drafting effectively of what a game might be. But as of today, AI is not able to take over the mainstream production process or the mainstream game operation process…
…It actually will generate, I think, a new generation of AI-native game experiences. These game experiences can actually be standalone, but they can also be integrated into evergreen game platforms, right? I think we have seen a lot of cases in which, if you see an innovation, then the evergreen game platforms can integrate these game experiences into the platforms and benefit from that great idea. At the same time, the AI game generation tools can actually help to make the game platforms evolving into more and more of a user-generated content, right? It will be sort of both the first-party and third-party game platform, right? You would have the professionally generated content by the game team, but at the same time, you actually sort of have a place for the users to generate their own content, right? I think we have already seen that doing very well for Peacekeeper Elite, right? We expect with the advent of more and more of these tools, right, it would actually help to expedite the evolution of many of our evergreen games into evergreen game platforms, and that would be hugely positive for us.
Tencent’s management thinks search is a very important capability in the AI age; Tencent has developed Yuanbao Search’s capabilities to the extent that it can be offered as a service
In terms of Yuanbao Search, it’s really a recognition that search is actually a very important capability in the AI age. It’s actually completely revolutionized with AI agentic capability. When we keep on improving the capability of Yuanbao, the search capability within Yuanbao becomes stronger and stronger, up to a point that we actually felt as an atomic capability, it actually can be offered to other internet and technology companies to generate revenue. That’s the reason why we actually sort of open it up and start offering it as a service and providing it to other companies. It’s a reflection of, one, AI has been transforming the search capability.
Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet (parent of Google), Amazon (parent of AWS), Meta Platforms, Microsoft, Nu Holdings, Sea, and Tencent. Holdings are subject to change at any time.