Nvidia Targets AI PCs With New Blackwell Chip and MediaTek CPU

Nvidia used Computex to put a new target in view: the personal computer. Jensen Huang introduced RTX Spark as a single-chip system for AI PCs, pairing a Blackwell RTX GPU with a custom Grace CPU built with MediaTek. The specifications shown in Nvidia’s presentation were meant to signal that this was not just another gaming GPU inside a PC: 6,144 CUDA cores, one petaflop of FP4 AI performance, a 20-core Grace CPU, NVLink C2C, and 128 gigabytes of LPDDR5X unified memory with 600 GB/s GPU-to-CPU bandwidth.

Ian King said Nvidia’s entry into the CPU side of the PC market had been rumored for a long time and was now being framed around a specific premise: if AI moves onto laptops, and users expect those machines to “think,” the devices will need a different kind of chip. Huang’s pitch, in King’s telling, was essentially: here is one prepared for that shift.

The market reaction treated the move as consequential. Nvidia was shown up about 4% during the program, while Arm was shown up 15.49% in an intraday market table. Ed Ludlow framed Nvidia’s CPU push, based on Arm architecture, as a challenge to the traditional PC territory of Intel and AMD. King and Ludlow described pressure on Intel and AMD from Nvidia’s move, though the visible intraday table in the source showed both Intel and AMD in positive territory at that moment.

King emphasized that Nvidia had not provided Bloomberg with a direct performance comparison against existing x86 or Arm PC offerings. Asked repeatedly what the chip does that others cannot, Nvidia’s answer was to wait for the devices. That left the market underwriting the credibility of Nvidia itself. King noted that attempts to bring Arm architecture into PCs have often struggled, with Apple as the arguable exception. Nvidia, however, has the resources, the PC brand through gaming GPUs, and a technology narrative tied to local AI.

He also cautioned against treating the concept as fundamentally new. AI PCs already exist, including Qualcomm-based systems that Ludlow said he and King had tested. King’s answer was that Arm in the PC “is not there” in a broad sense and that Nvidia itself had tried a related effort more than a decade earlier without success.

Bloomberg Intelligence’s framing, shown on screen, said Nvidia’s RTX Spark family is initially aimed at developers and enthusiasts running models locally, while also arguing that Mac Mini demand suggests the category could scale over time. Mandeep Singh made the broader analogy to servers: servers had been a low- to mid-single-digit growth market before AI servers changed the category. Nvidia, now a much more recognizable brand than it was a few years ago, is testing whether a comparable reframing can work in PCs.

Singh did not present the PC market as broadly healthy. Ludlow noted that IDC had a forecast for the overall PC market to decline 11% in “calendar 26,” alongside memory-pricing pressure and broader hardware-cycle issues. Singh’s point was narrower: the AI PC category can diverge from the aggregate PC market if agentic AI requires much more compute across the board. He referred separately to 132 gigabytes of DRAM in discussing the new class of machines, while the Nvidia slide shown in the source listed 128 gigabytes of unified memory. In his view, Nvidia is trying to build an ecosystem that spans physical AI, client PCs, and the data-center stack.

Jensen Huang used Computex to argue that the next phase of AI adoption is agentic. Bloomberg’s Stephen Engle reported from Taipei that Huang called agentic AI “useful AI,” described the current period as the “agentic age,” and said Nvidia’s Vera Rubin AI supercomputing platform is now in full production. The same keynote that pushed Nvidia into the PC also pushed a broader claim: if billions of AI agents become active users of software, software companies may see more demand rather than less.

That claim showed up immediately in the market segment. ServiceNow, Adobe, and Salesforce were shown rallying sharply after Huang’s remarks. On-screen intraday data showed ServiceNow up roughly 10%, Adobe nearly 6%, and Salesforce about 10%. Ludlow summarized Huang’s thesis this way: billions of AI agents could become the biggest users — and effectively the biggest customers — of software companies.

Huang’s job-market claim was one of the strongest lines from the Computex remarks. He argued that AI will be a generator of GDP rather than a destroyer of work. Singh was receptive to the software-demand argument, but not without limits. He said recent results from companies including Datadog, Snowflake, and MongoDB made it hard to support a thesis in which software is fully disintermediated and rebuilt from scratch by agentic AI. The user interface may become more conversational and natural-language-driven, but systems of record and enterprise plumbing built over years are not simply disappearing. Singh said some of these companies had shown that in their numbers, with ARR growth accelerating.

Singh also accepted that agentic AI could create a multiplier effect for technical knowledge workers: one worker using 10 or even 100 agents. But he did not treat the employment impact as settled. College graduates were trained for a particular career progression, he said, and that model has already changed and will change further.

MediaTek’s Vince Hu echoed Huang from the partner side. Hu said MediaTek is not reducing engineering headcount; it is expanding and hiring engineers as AI increases momentum across automotive, IoT, data center, wearables, and existing businesses. He also argued that Taiwan may be in a technology “heyday,” citing strong GDP growth forecasts discussed by Bloomberg’s Stephen Engle and saying MediaTek is working with tier-one hyperscalers. Hu said he felt good about the fundamentals of the business “at least through the 2030 horizon.”

The investment implication, for Matt Wittmer, was not to treat Nvidia as the whole trade. Wittmer said investors still have reasons to hold large-cap technology names because of cash-flow visibility and positioning in AI, especially during a market backdrop that included Middle East geopolitical risk. But his central message was diversification. Concentration risk, he said, is real, and AI capital spending is beginning to spread across what he called the AI value chain.

Wittmer divided that chain into five layers: energy and power; chips and computing infrastructure; hyperscalers; AI models; and the application layer. His emphasis was on the top of the stack. He argued that software and applications are where industries will pursue automation, decision support, and productivity gains. Companies with access to important data, operational scale, and deep embedding in customer workflows are the ones he expects to matter as enterprises adopt AI.

Wittmer also identified memory as one bottleneck he does worry about. He said clean-room capacity is limited and greenfield projects can take two to three years to come online. That supply constraint, in his view, supports continued pricing power for memory names. His investment answer was not to abandon Nvidia or infrastructure, but to avoid treating any one layer as the whole AI trade.

For Amit Jain, the hard problem in robotics is not whether a robot can be trained to perform a single task. It is whether robots can generalize. Today, he said, most robots are trained by showing them examples of specific tasks, one task at a time. That contrasts with large language models, which are trained broadly and can handle a wide range of unseen problems.

Jain defined generalization in robotics as the ability to solve generally any task, including a new scenario encountered for the first time. Without that, robots remain stuck in what he called “the valley of specific tasks.” The desired endpoint is a system users can speak to naturally: do this, then when finished go handle that, or run the house.

Luma AI is launching an open research lab focused on physical AI, and Jain framed openness as both technical and political. The brute-force approach to robotics, in his account, would be to gather data on every task and every combination of tasks humans perform: picking up cups, digging mines, and everything else. He called that practically impossible. Luma’s alternative is to use large-scale multimodal infrastructure and internet-scale multimodal data to extract signals that support control, simulation of reality, and physical control.

Jain said Luma’s work over the past four years has been building general systems from multimodal data, including 3D, images, and video from raw internet data. That, he argued, is the relevant capability for physical AI. He contrasted it with a domain-expert view of the problem, saying language models were not solved by linguists and physical AI will require large-scale multimodal data systems.

The open-science position was more than branding. Jain argued that physical AI systems will be present in homes, manufacturing, food production, hospitals, scientific labs, and streets. Because they could control or participate in the means of production, he said it is “completely untenable” for one or two actors to control the full stack. Caroline Hyde pressed him on whether that was a philosophical position and whether it can survive the economics of frontier AI, the retreat of some companies from open source, and geopolitical concerns around China. Jain answered that it has to be philosophical because physical AI is not just a tool, and he also argued it is economically sound: nations will not accept a future in which one or two foreign companies control their means of production.

His proposed model is an ecosystem of chip partners, model providers such as Luma, and deployment partners building productive systems together. Jain also said current large language models are moving in the “wrong path” on this question of control.

SpaceX’s prospective IPO was presented as a market-structure event before it is a normal listing. The company is pitching investors on a claimed market opportunity of $28.5 trillion, and Bloomberg’s reporting framed the scale as large enough to challenge assumptions about valuation, index inclusion, and passive demand.

Isabelle Lee said index providers have already begun revising their rules. Nasdaq has shortened the waiting period before a newly public company can be considered for the Nasdaq 100 from three months to 15 days. FTSE’s period is five days. S&P, she said, is consulting on moving from 12 months to six months. Supporters argue that faster inclusion better reflects the market. Critics argue that newly public companies are often volatile and that investors should allow more time for price discovery.

The passive-investing consequence is large. Lee said index-tracking funds are estimated to buy nearly $20 billion worth of SpaceX. Hyde framed the IPO as including a 30% allocation to retail investors, and Lee described the resulting tension: some retail traders may ask why they should buy in the IPO if they will soon own SpaceX indirectly through index funds. The question becomes not merely whether SpaceX is attractive, but how quickly it becomes unavoidable in ordinary brokerage accounts.

Lee said the concentration issue goes beyond SpaceX. A bet on Tesla and a bet on SpaceX can both be read, in market terms, as a belief in Elon Musk. Even though the companies are separate and contain different businesses, she said “it’s all really Elon at the end of the day.” If SpaceX is added to indexes, investors already exposed to Tesla may have to ask whether they are overconcentrated in one person’s companies. She extended the issue to future listings from companies such as OpenAI and Anthropic: ordinary portfolios could become even more concentrated in technology.

George Ferguson later walked through Bloomberg Intelligence’s sum-of-the-parts framework, which suggested comparables could support a SpaceX valuation near $2 trillion. He was careful not to present that as a definitive fair value. The numbers, he said, “seem pretty rich.” The largest component was the launch business, valued at slightly over $1 trillion in the displayed analysis. Starlink added about $600 billion, and xAI was estimated at $300 billion to $400 billion.

Ferguson said the launch-business estimate came from market comparables, especially Rocket Lab, which he said trades at about 80 to 90 times revenue. For SpaceX, Bloomberg Intelligence added internal launch activity that does not appear in the revenue line, arriving at about $11 billion of launch revenue before applying a comparable multiple. Ludlow challenged the launch number because Musk has said in the past that launch revenue has a ceiling, which is why the company’s total addressable market pitch reaches far beyond putting payloads into orbit. Ferguson’s answer was that the analysis was comparable-driven, not a statement that the price was unquestionably justified.

Hyde also raised a separate IPO-mechanics question, citing Bloomberg’s story that SpaceX was going to reserve up to 5% of IPO stock for certain employees and friends and asking about the issue if those holders were not bound to a lockup in the same way. Ferguson did not confirm the structure in detail. He answered more generally that IPOs often face pressure after the initial hype fades. SpaceX’s scale and mix of AI, launch, and connectivity may lessen that hangover, he said, but he still expected challenges.

Mark Gurman argued that Apple’s expected move into smart glasses should be understood through the Apple Watch playbook rather than through Vision Pro. When the Apple Watch launched in 2015, he said, it upended the mid-tier mechanical watch industry within a few years. Bloomberg’s on-screen comparison showed Swatch’s 2025 revenue 28% lower than in 2014 and Fossil’s 2025 sales 70% lower than in 2014.

Gurman’s distinction was between the middle and the high end. Companies such as Movado, Fossil, and Swatch were pressured, while luxury watchmakers such as Patek, Rolex, and Audemars Piguet did fine. He expects a similar pattern in eyewear: smart glasses from Apple and others could disrupt the mid-tier market, including brands such as Ray-Ban and Warby Parker, while very high-end brands such as Maison Bonnet and Cartier may remain insulated.

The market is already shifting. Gurman said Ray-Ban and Warby Parker have pivoted toward smart glasses, and Meta has done a strong job pioneering the category through its partnership with EssilorLuxottica. He expects Meta to launch new smart glasses and eventually introduce more in-house form factors. Apple’s likely entry, in his view, would increase interest across the category, with Apple serving iPhone users and Meta having a strong chance on the Android side.

Vision Pro, Gurman said, is a separate story. It is a $3,500, heavy, “tech tour de force” device. Smart glasses would be lighter and much cheaper. The comparison is not that the products are similar, but that Gurman expects Apple to enter an adjacent consumer category and change expectations for the mainstream version of it.

New York Tech Week opened with what Julie Samuels described as a booming local sector. She said the city is hiring in tech at twice the rate of San Francisco and four times the rate of Boston, has more than 9,000 startups by the last count she cited, and has seen venture funding rise sharply year over year. A Tech:NYC graphic shown during the segment described New York as the second-largest and most valuable tech ecosystem globally and said first-quarter tech funding surged 172% year over year to $11 billion.

Samuels’ argument was that AI companies increasingly need to be close to the sectors they affect. If the customers are in finance, health care, media, enterprise services, and other major industries, then New York offers expertise, smart capital, customers, clients, mentors, and network density. She said large platforms often build technology on the West Coast and then come to New York to answer commercialization questions: who buys it, how it is monetized, and how it is used.

She said the pattern has intensified over the past year. Frontier labs including OpenAI and Anthropic are hiring in New York and doing large real estate deals, while startups are also coming to the city. The startup profile, in her view, is different from San Francisco’s. San Francisco produces more “really, really hard tech,” while New York startups are often more integrated into existing sectors. Samuels called that healthy for both the city and the country.

Asked about the pitch to highly paid tech workers who have options, Samuels leaned on urbanism. New York’s density, public transit, diversity, and cultural energy make it attractive to builders, she said. She also contrasted the social texture of the two cities: in San Francisco, one meets tech people who live in San Francisco; in New York, one meets New Yorkers who work in tech. Her point was not that one replaces the other, but that both dynamics matter.