NVIDIA RTX Spark and Computex 2026 Keynote

RTX Spark's headline number -- 1 petaflop of local FP4 AI compute -- only makes sense once you see how the die is wired. A 20-core Arm Grace CPU sits next to a Blackwell RTX GPU with 6,144 CUDA cores and 5th-generation FP4 Tensor Cores, and the two are stitched together with NVLink-C2C across 128GB of unified memory ^[1]. That last detail is the trick. On a normal laptop, model weights have to be shuttled between CPU RAM and GPU VRAM over a relatively narrow bus, which is what makes running anything larger than ~13B parameters miserable on the road. With unified memory and a chip-to-chip interconnect, the GPU can address the full 128GB pool directly, and a model never has to leave.

The practical envelope NVIDIA quotes follows from that: render ultralarge 90GB+ 3D scenes, edit 12K 4:2:2 video, generate 4K AI video, run 120-billion-parameter LLMs with up to a 1-million-token context window using local agents, and still play AAA games at 1440p over 100 FPS ^[2]. Translation for a builder: the same 14mm, 3-pound laptop that runs an indie game at high refresh can also host a frontier-class assistant that ingests an entire codebase, an entire deposition, or an entire video edit without round-tripping to a cloud API. The chip is manufactured on TSMC's 3nm EUV node ^[3], and Huang committed to a three-generation Spark roadmap -- Grace Blackwell now, Vera Rubin with LPDDR6 memory next, then Rosa Feynman ^[4]-- signalling this is a platform, not a one-off.

Wall Street treated the keynote as a zero-sum event. Within hours of Huang walking off stage, Intel fell ~6%, AMD ~5% and Qualcomm shed more than 6% -- the three companies whose Windows volumes Spark most plausibly steals from ^[5]. On the other side of the trade, Dell rose 10%+ and HP ~8% as Morgan Stanley flagged that Dell's 'better access to memory supply and pricing than many enterprise peers' is a structural edge for shipping 128GB unified-memory machines at scale ^[6]. Microsoft itself was up about 3% ^[7].

The wildest move was Marvell, which surged roughly 19-22% pre-market after Huang's on-stage shoutout calling it the 'next trillion-dollar company' ^[8]. That single line moved more market cap than most full earnings reports. Read together, the tape says investors don't see this as additive -- they see a finite Windows premium-laptop pie being re-sliced toward Arm + NVIDIA, and they're willing to fund the OEMs (Dell, HP) closest to the new silicon while marking down the incumbents most exposed to it. Even Intel's own response -- admitting a 'healthy dose of paranoia' while touting the virtues of x86 ^[9]-- is unusually frank for a vendor that historically dismisses Arm threats.

Spark didn't appear because NVIDIA suddenly figured out PC silicon -- it failed at that publicly with Tegra in the early 2010s ^[10]. It appeared because three things converged in 2026. First, Microsoft's Windows-on-Arm program had been an effective Qualcomm exclusive since 2018, locking out other chipmakers for roughly eight years; that deal lapsed, opening the slot RTX Spark just filled ^[10]. Second, MediaTek -- best known for smartphone SoCs -- designed the custom 20-core Arm CPU complex inside Spark, formally announced one day before the keynote on May 31, 2026 ^[11]. That partnership gave NVIDIA a CPU team without having to build one from scratch, and gave MediaTek its first serious foothold in Windows.

Third, Apple Silicon spent five years proving to the Windows world that a vertically integrated Arm SoC with unified memory beats discrete CPU+GPU+RAM on perf-per-watt for the workloads laptop buyers actually care about. RTX Spark is explicitly framed as the Windows-on-Arm answer to the Apple M5 family ^[12]. The pitch lands the way it does -- Nadella calling it a breakthrough toward 'unmetered intelligence to every home and every desk with Windows' ^[1], Huang calling it 'the first across-the-lineup PC reinvention in forty years' ^[13]-- only because the cultural and contractual conditions for an Arm Windows premium laptop finally exist at the same time.

The press coverage is glowing; the engineers running the numbers are not. On developer forums, the people who would actually buy Spark to run frontier models locally are doing the math and finding it doesn't quite close. The 128GB unified memory ceiling rules out the largest open-weight models people now expect to run at home -- GLM and Kimi-class systems -- and rumored top-SKU pricing in the $3,500 range is well above the $1,500 threshold DigiTimes analyst Jason Tsai says Spark needs to hit to break out of niche status ^[14]. The other technical complaint is memory bandwidth: even with NVLink-C2C, Spark is being compared unfavorably to the bandwidth of an Apple M5 Max, which matters more than peak FLOPs for inference on long contexts. Mainstream gamer threads are more eye-rolling than analytic -- the dominant reaction to the agentic-AI framing is skepticism that any of it ships polished -- and an investor-side contrarian read explicitly invokes a 'Tegra 2.0' pattern, betting that NVIDIA will once again discover that consumer PC margins and support burdens punish silicon outsiders.

Developer YouTube is split along similar lines. Long-form reviewers treat the early Spark hardware as a preview and reserve judgement, while the most-watched takedown angles directly at the unresolved Windows-on-Arm app-compatibility tax -- the same problem that quietly capped Qualcomm Snapdragon X adoption ^[15]. The investor case and the engineer case have not yet been reconciled: equity markets are pricing Spark as if compatibility, pricing, and bandwidth will be solved by Fall 2026, while the people closest to those exact problems are skeptical the solution arrives on time.