Nvidia Enters $200 Billion CPU Market with AI-Focused PCs via Microsoft, Dell, and HP

At Computex 2026 in Taipei, Nvidia CEO Jensen Huang unveiled RTX Spark, the company's first processor designed for Windows PCs. The chip combines a 20-core Grace CPU built on Arm architecture with a Blackwell RTX GPU packing 6,144 CUDA cores, connected through NVLink-C2C at 600 GB/s [1]. Nvidia claims 1 petaflop of AI compute — roughly 1,000 TOPS (trillion operations per second) — in a package thin enough for 14mm laptops weighing three pounds [1].

Systems running RTX Spark will ship this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and Gigabyte to follow [1]. The chip was co-developed with MediaTek, which contributed the CPU architecture, power management, and a memory controller supporting up to 128 GB of unified LPDDR5X [2]. Microsoft CEO Satya Nadella called RTX Spark "a real breakthrough" for delivering AI "to every home and desk" [1].

The pitch is ambitious: Nvidia wants the PC itself to become an AI agent host, running 120-billion-parameter large language models locally with up to 1 million tokens of context [1]. But to get there, Nvidia must break into a market where Intel and AMD collectively control more than 95% of processor volume [3], where the last Arm entrant captured less than 1% share after a year on shelves [4], and where enterprise buyers replace hardware on four- to five-year cycles regardless of what's announced at trade shows [5].

The Market Nvidia Wants to Crack

The global PC CPU market generates roughly $200 billion in annual revenue [6]. Intel and AMD have split it between them for decades, and the latest Mercury Research data from Q4 2025 shows how lopsided that split remains: Intel holds 63.6% of desktop and 74% of laptop CPU unit share, with AMD at 36.4% and 26% respectively [3]. Qualcomm's Arm-based Snapdragon X, the most recent architecture to challenge x86 in consumer PCs, managed just 0.8% of the total market [4].

Source: Mercury Research / Tom's Hardware

Data as of Mar 1, 2026CSV

Nvidia currently has zero share of the consumer PC processor market. Its Grace CPU has shipped exclusively in data center configurations, where benchmarks showed the 72-core Arm chip trading wins with AMD's EPYC and Intel's Xeon in server workloads [7]. The RTX Spark represents a fundamentally different product: a system-on-chip (SoC) combining that server-class CPU lineage with Nvidia's GPU, targeting laptops and compact desktops rather than rack-mounted servers.

The Performance Argument: 1,000 TOPS vs. 50

The core of Nvidia's sales pitch is a raw performance gap in AI workloads that makes existing "AI PCs" from Intel and AMD look like toys. Intel's Lunar Lake NPU delivers 48 TOPS, AMD's Strix Point offers 50 TOPS, and Qualcomm's Snapdragon X Elite provides 45 TOPS [8][9]. RTX Spark claims 1,000 TOPS — a 20x multiple [1].

Source: Manufacturer specifications

Data as of Jun 1, 2026CSV

That comparison, however, requires context. The TOPS figures for Intel, AMD, and Qualcomm measure only the NPU (Neural Processing Unit), a dedicated accelerator handling specific AI tasks like image recognition or noise suppression. Nvidia's 1,000 TOPS figure encompasses the entire SoC, driven primarily by the Blackwell GPU's Tensor Cores operating at FP4 precision [1]. The NPU in a Lunar Lake chip was never designed to run a 120-billion-parameter language model — it handles smaller, targeted inference tasks within a strict 15-watt thermal envelope [8].

The question is whether mainstream buyers need GPU-class AI compute. Intel's Lunar Lake achieves 100+ TOPS when combining CPU, GPU, and NPU, which is sufficient for Microsoft's Copilot+ features and the vast majority of on-device AI tasks shipping today [8]. Most general-purpose AI workloads sustain 15–25 TOPS in real-world operation [8]. Running a full LLM locally is technically impressive but remains a niche use case for developers, researchers, and AI enthusiasts — not the 250 million PCs that ship each year to office workers and students.

The Qualcomm Precedent: A Warning Sign

Nvidia's OEM roster — Dell, HP, Lenovo, Microsoft, ASUS, MSI — looks nearly identical to the partner list that accompanied Qualcomm's Snapdragon X Elite launch in mid-2024. That launch arrived with similar enthusiasm about Arm efficiency, AI acceleration, and a new era of Windows computing. The results were sobering.

Snapdragon X shipped approximately 720,000 units in Q3 2024, its first full quarter, capturing just 0.8% of the total PC market [4]. By December 2024, Qualcomm claimed 10% of U.S. Windows PC retail sales at $800 and above — but that metric excluded institutional purchases (which represent the majority of Dell and HP commercial volume) and counted only premium-priced devices [10]. Qualcomm has since announced 80+ PC designs for 2025 and 100+ for 2026, but market share gains have been incremental [10].

The parallel is instructive. Qualcomm offered meaningful advantages in battery life and had full backing from Microsoft and major OEMs. It still hit the same wall: x86 application compatibility. Windows on Arm runs x86 and x64 software through Microsoft's Prism emulation layer, but kernel-level drivers — including some enterprise antivirus products, VPN clients, and security software — require native Arm builds [11]. Anti-cheat software used in gaming blocks execution on Arm hardware entirely [11]. For enterprise IT departments that manage thousands of devices running specialized line-of-business applications, the risk of a single incompatible driver or plugin can disqualify an entire architecture from procurement.

Nvidia and its partners are betting that RTX Spark's superior GPU performance will make the compatibility tradeoff worthwhile in ways Qualcomm's efficiency advantage did not. Windows Central's analysis of the launch suggested RTX Spark is "less of a threat to Qualcomm's chips and more of an ally to Microsoft's Windows on Arm PCs" — expanding the Arm ecosystem rather than directly displacing Qualcomm [12]. Whether a larger Arm ecosystem accelerates native software development or merely spreads the same compatibility problem across more hardware remains an open question.

Intel and AMD's Defense: x86 Isn't Standing Still

Intel's response to RTX Spark has been measured but pointed. At Computex, Intel acknowledged having "a healthy dose of paranoia" about Nvidia's entry while calling it "great for the market" [13]. The subtext: Intel believes x86's software compatibility moat — every Windows application, driver, and enterprise tool built over 40 years — is wide enough to withstand a performance gap in AI-specific workloads.

Both Intel and AMD have shipped processors with NPUs rated above 40 TOPS, meeting Microsoft's Copilot+ PC requirements [8][9]. AMD's desktop CPU revenue share reached 42.6% in Q4 2025, its highest ever, driven by competitive pricing and strong per-core performance [3]. Intel's Panther Lake, expected later in 2026, will increase NPU capacity further while maintaining full x86 backward compatibility.

The steelman case for x86 incumbents rests on a simple question: what concrete consumer workload requires 1,000 TOPS that cannot run adequately on 50 TOPS? Microsoft's own Copilot+ features — live captions, image generation, intelligent search — work within the 40 TOPS envelope. Adobe is rearchitecting Photoshop and Premiere for RTX Spark with promised 2x speedups [1], but those applications already run natively on x86 with GPU acceleration through CUDA on discrete Nvidia graphics cards that millions of creative professionals already own.

Running 120-billion-parameter LLMs locally with 128 GB of unified memory is genuinely new capability for a laptop form factor. But until software ecosystems produce compelling consumer applications that require that capability — not just tech demos — Intel and AMD can argue their existing platforms handle every shipping AI feature at lower cost and with zero compatibility risk.

The AI PC Market: Big Numbers, Uncertain Demand

The broader AI PC market is growing rapidly on paper. Gartner projects AI PCs will account for 31% of global shipments in 2025 (~77.8 million units) and 55% by 2026 [14]. Counterpoint Research forecasts AI Advanced PC penetration reaching 59% of global shipments in 2026, up from 39% in 2025 [15].

Source: Gartner / Counterpoint Research

Data as of Jun 1, 2026CSV

These numbers require careful interpretation. The definition of "AI PC" has been elastic — any machine with a dedicated NPU qualifies under most industry definitions, meaning every new Intel, AMD, and Qualcomm laptop already counts. The 100 million AI PCs projected for 2025 are overwhelmingly x86 machines running Intel and AMD silicon [14]. Nvidia's RTX Spark won't ship until fall 2026 and will initially target premium segments, meaning its contribution to these totals will be minimal in the near term.

The more relevant question is unit volume at the scale needed to attract third-party developer investment. Given that the average enterprise PC replacement cycle runs 3–5 years [5] and consumer replacement cycles have stretched to 5–6 years [5], the installed base of RTX Spark machines will grow slowly. Software developers port applications to new architectures when the addressable device count justifies the engineering cost. Apple's M-series transition succeeded in part because Apple controls both hardware and software, compelling developers to adapt. Nvidia controls neither the operating system (Microsoft does) nor the application ecosystem.

Who Bears the Risk if AI PCs Underdeliver

Dell and HP know what happens when PC demand projections miss. In 2023, global PC shipments fell 14.8% to 241.8 million units — the worst year in industry history — following a post-pandemic demand collapse [16]. Dell shipped 40.2 million units, down 19.5% from 2022; HP shipped 52.9 million, down 4.5% [16]. Dell's inventory fell to $4.78 billion, a 19% decline, as the company worked through excess stock [16].

Those losses inform how aggressively OEMs will commit to an unproven architecture. Specific contractual terms between Nvidia and its OEM partners have not been disclosed. But the pattern in PC manufacturing is well established: OEMs typically make volume purchase commitments months in advance of product launches, and when demand disappoints, they absorb inventory write-downs. The 2023 downturn cost billions across the industry in markdowns and unsold stock.

Nvidia's structure mitigates its own downside risk. As a chip designer using TSMC for fabrication and MediaTek for CPU design collaboration [2], Nvidia's capital expenditure exposure is lower than an integrated manufacturer's. The OEMs ordering RTX Spark systems bear the primary inventory risk if consumer and enterprise demand fails to materialize at projected volumes.

Antitrust Exposure: Vertical Integration Under Scrutiny

Nvidia's move into PCs extends an already sprawling vertical integration strategy. The company now controls GPU silicon, CPU silicon (Grace/Vera), networking (Mellanox/InfiniBand), and the dominant AI software stack (CUDA). Adding PC processors means Nvidia would supply GPU, CPU, and software platform to the same OEMs — a bundling arrangement that regulators have previously scrutinized.

The European Commission has already sent questionnaires to Nvidia customers investigating whether the company engages in commercial or technical tying of GPU products — specifically, whether contracts require customers to purchase networking equipment alongside GPUs, or whether discounts incentivize buying Nvidia hardware and software as a package [17]. EU antitrust violations carry fines of up to 10% of global annual revenue [17].

No regulator has publicly opened an inquiry specifically targeting Nvidia's PC CPU entry or any exclusivity arrangements with the named OEM partners. But Nvidia's 84% share of the AI accelerator market already makes it a regulatory target [17]. Extending that dominance into the PC processor market — potentially bundling CUDA software optimization, GeForce GPU drivers, and CPU silicon into a vertically integrated platform — will attract attention from competition authorities in the EU, US, and China, all of which have shown increasing willingness to intervene in semiconductor market concentration.

The Upgrade Timeline: Math That Favors Incumbents

Even under optimistic assumptions, Nvidia faces an arithmetic problem. With enterprise refresh cycles of 3–5 years and consumer cycles of 5–6 years [5], the installed base of PCs turns over slowly. Windows 10 end-of-support in October 2025 is accelerating some upgrades [18], but those purchases are primarily landing on Intel and AMD hardware shipping now — not on RTX Spark systems that won't reach shelves until fall 2026.

If RTX Spark captures even 5% of new PC shipments in its first full year (roughly 13 million units out of ~260 million total), that would represent a strong launch by historical standards — Qualcomm hasn't reached that level after two years [4]. But 13 million units is a fraction of the installed base Intel and AMD service, and likely insufficient to motivate widespread native Arm software porting beyond what Qualcomm's presence has already prompted.

For Nvidia to reach meaningful scale by 2027 — say, 30–40 million cumulative units — it would need aggressive OEM pricing, rapid enterprise adoption despite compatibility concerns, and sustained consumer demand for local AI workloads that don't yet have clearly defined use cases beyond running LLMs. Each of those conditions is individually plausible; requiring all three simultaneously is a high bar.

What This Actually Changes

RTX Spark is Nvidia's boldest strategic bet outside the data center. The chip's raw AI performance is genuinely unprecedented for a laptop form factor. The OEM and software partnerships are real. The MediaTek collaboration provides credible CPU engineering. And the broader trend toward on-device AI computation is durable.

But the history of PC architecture transitions — from PowerPC to Intel in Apple's case, from x86 to Arm in Qualcomm's — suggests that superior silicon alone doesn't win markets. Software compatibility, enterprise procurement inertia, replacement cycle timing, and developer ecosystem investment all matter as much as TOPS. Nvidia's 1,000 TOPS advantage over Intel's 48 is dramatic on a spec sheet. Whether it translates into the 50–100 million unit installed base needed to sustain a self-reinforcing platform is a question that won't be answered until well into 2028.

Jensen Huang framed the bet succinctly at Computex: "The PC is being reinvented...you ask — and the PC does the work" [1]. Intel, AMD, and 40 years of x86 compatibility are the work RTX Spark must now undo.