AMD Buys Taalas, Nvidia Trims Rubin Ultra: Aug. 7

AMD made two notable moves in 48 hours: it beat Q2 earnings and acquired a startup that bakes AI models directly into silicon. Meanwhile, Nvidia is quietly reworking its next flagship GPU because memory suppliers cannot keep up.

Key points

• AMD acquired Toronto-based Taalas, a startup whose chips hardwire AI inference models into silicon to cut compute and memory bottlenecks.
• Taalas currently runs a compact version of Meta’s Llama 3.1 on its chip, with more advanced models in development.
• AMD also beat Q2 earnings and raised forward targets, with data center momentum cited as the primary driver.
• Nvidia is testing lower-memory configurations of its Rubin Ultra GPU as a worsening high-bandwidth memory (HBM) shortage pressures the 2027 product roadmap.
• SpaceX’s IPO lockup expired Thursday, unlocking more than 911 million shares for early investors, while space-sector peers including Rocket Lab and ASTS have rebounded 28% to 54% from their July lows.

What does the Taalas deal tell us about AMD’s inference strategy?

AMD is making a pointed bet that the next phase of AI competition runs through inference efficiency, not just raw training throughput. Taalas develops specialized silicon designed to reduce the compute and memory bottlenecks that slow down deployed AI models. Hardwiring a specific model architecture into silicon can cut latency and power consumption sharply compared to a general-purpose GPU running the same workload.

The acquisition fits a broader pattern. As AI inference volumes scale, the economics favor purpose-built silicon over flexible but power-hungry GPUs. AMD’s Q2 beat and raised guidance suggest data center revenue is already strengthening. Adding Taalas gives AMD a foothold in the edge and on-device inference market, where Qualcomm and a handful of startups have been moving aggressively. The deal also adds competitive texture against Nvidia, whose dominance in training has always been less clear-cut on the inference side.

Taalas is small and its current chip targets a modest model. The commercial impact will depend on how quickly AMD can integrate the technology into products at scale, and whether the architecture extends beyond Llama-class models.

Why is Nvidia testing a leaner Rubin Ultra?

According to Digitimes, Nvidia is evaluating significantly lower-memory variants of Rubin Ultra, its next-generation GPU, in response to a tightening HBM supply picture for 2027. This is a meaningful signal. Rubin Ultra is supposed to be Nvidia’s flagship product for the next infrastructure cycle, and redesigning memory configurations at the testing stage suggests supply constraints are serious enough to reshape specs rather than just limit volumes.

The downstream effects could be considerable. Cloud providers and AI developers currently size their GPU clusters partly around memory capacity per chip. Leaner HBM configurations could force those buyers to model out more GPUs per workload, adding cost and complexity. At the same time, it hands HBM suppliers including SK Hynix and Samsung additional pricing leverage heading into contract negotiations for 2027 builds. Investors in memory stocks will be watching whether that leverage materializes in margins.

Microsoft’s Azure acceleration and the broader cloud read-through

Separate from the chip moves, Microsoft’s Q4 results showed Azure acceleration and early signs of Copilot monetization, according to analysis published today. That matters for chip investors because Azure growth is a direct demand signal for GPU infrastructure. A durable Azure acceleration, if it holds through the second half of 2026, would argue against the AI spending skepticism that knocked Marvell down 37% in July and contributed to a record 15.2% monthly decline among Asian hedge funds focused on AI-linked equities.

SpaceX lockup: what happens next?

SpaceX’s IPO lockup expired Thursday with over 911 million shares now eligible for sale. Options traders reportedly see signs the stock may be approaching a floor, but the sheer size of the unlocked float creates near-term overhang. Direxion responded to market appetite by launching a 2x inverse SpaceX ETF (LOFD), underscoring how much short-side interest has built up. Meanwhile, sector peers Rocket Lab and ASTS have recovered sharply since early August, suggesting some capital that rotated out during SpaceX’s IPO-driven dip is finding its way back into the broader space equity complex.

Nothing here is investment advice. All figures sourced from the articles linked above.