Macro_Lin
Semiconductor Analyst, AI Chip Design, Ex-Quant/Portfolio Manager
Nvidia's scaled-back Rubin Ultra is not good news. Per @SemiAnalysis_, just three months after it was announced at GTC 2026, the original four-die Rubin Ultra has been cancelled over manufacturing execution concerns, and the new version is half the size with roughly half the real-world performance. x.com
Licensing Groq's inference technology lets Nvidia push high-performance inference chips into the large, long-running inference market, which should drive a re-rating. Per Reuters, Nvidia plans a China-sellable Groq chip around May that is neither cut down nor China-specific; Groq's SRAM-based LPU, licensed for $17bn, gives very low latency and energy use. Wall Street had largely written off China revenue; rivals like Baidu and Cambricon may lose their window if Beijing does not block it. x.com x.com
Zephyr
AI & Semiconductor Analyst at Citrini Research
Nvidia also reduced the NVLink bandwidth per GPU from 1.8 TB/s (3.6TB/s bidi) to 1.5 TB/s (3 TB/s bidi) x.com
AlvaSupportsThe reported 17% cut in Nvidia's per-GPU NVLink bandwidth adds a concrete interconnect downgrade to check against Rubin Ultra's projected system performance.
Damnang
Semiconductor Engineer in Silicon Valley, AI Chip Investment Research
One correction. My understanding is that both Vera Rubin and Vera Rubin Ultra are still being evaluated with a mix of 8H and 12H configurations, with both options kept open under an open lab approach. x.com
AlvaChallengesDamnang's understanding corrects the reported HBM configuration path: both 8-high and 12-high remain under evaluation for Rubin Ultra, leaving its memory design less settled than Daishin's staged forecast suggests.
Jukan
Semiconductor & AI Infrastructure Research at Citrini
NVIDIA has officially lowered the HBM4 pin speed. 11.7-> 9.3Gbps x.com
AlvaSupportsJukan reports Nvidia lowered HBM4 pin speed from 11.7 to 9.3 Gbps, a roughly 21% reduction in per-pin bandwidth that raises a concrete performance question for Rubin Ultra.Jukan
Semiconductor & AI Infrastructure Research at Citrini
NVIDIA: Rubin’s specifications have changed amid base die supply constraints and requests from customers prioritizing bandwidth over cost efficiency. ... For Rubin Ultra, scheduled for release in 2027, multiple architectural variants are currently undergoing tape-out. Initial versions are expected to use the highest-specification 8-high HBM4, while subsequent versions are expected to adopt either 8-high or 12-high HBM4E. x.com
AlvaMixedDaishin Securities reports base-die constraints and customer bandwidth preferences behind Rubin specification changes, while Rubin Ultra still has multiple variants in tape-out; this points to both supply pressure and an evolving design rather than a settled replacement.
Aravind Srinivas
Co-Founder & CEO of Perplexity AI
Tokens are going to grow like multifold, but the spend per token is not going to grow the same level. ... And whatever it costs to serve the models should also go down as you start post-training more cost efficient models. podtrac.com
AlvaSupportsHis expectation of multifold token growth supports the thesis's long inference-demand runway, while falling per-token spend and serving costs make the revenue upside less direct.
Dylan Patel
Founder & Chief Analyst of SemiAnalysis, AI Infrastructure Research
[Replying to @rev_cap: "Semianalysis saying if you’re worried about demand (ARR,) don’t be, because there’s going to be even more supply than you realize"] @rev_cap Demand and ARR is what to worry about, not supply. x.com
AlvaChallengesPatel identifies demand and ARR as the risks to watch, challenging the thesis's reliance on a long runway for inference demand.
Andrew Feldman
Co-Founder & CEO of Cerebras
We both agreed - people keep underestimating the demand for faster AI. There is no ceiling. 10x, 100x, 1000x. ... For 70 years, we have delivered more compute. And builders have used all of it. Every time. x.com
AlvaSupportsThis supports the thesis that expanding inference capacity can meet a long runway of demand: Feldman argues that faster AI will drive greater use of compute, with each gain spent on bigger problems.
Aaron Levie
Co-Founder & CEO of Box
Agent swarms are the first real ability to parallelize this work in a way where - in theory - the only parameter that matters is how much compute you throw at the problem. ... Cyber security, life sciences or material science research, trading, QA testing, data processing, and many other fields are essentially compute constrained problems in the limit once models are good enough at these tasks. x.com
AlvaSupportsThis supports the thesis that inference demand has a long runway: Levie argues that capable, coordinated agents could make compute a key limiting input across many economically valuable fields. The scale of that demand remains conditional on models becoming good enough.
SemiAnalysis
Independent Semiconductor and AI Research Firm
@Inferact & Google submit their FIRST preliminary TPUv7 Kimi K3 2.8T Model Submission, and it is already delivering strong perf per TCO against GB200 NVL72, B300, and MI355 in the ultra-high-interactivity regime due to their megakernel approach for external customers. x.com
AlvaChallengesThis weakens the inference-demand upside for Nvidia by pointing to a competing TPU platform with strong preliminary performance per TCO against Nvidia systems in ultra-high-interactivity workloads.
Harrison Chase
Co-Founder and CEO of LangChain
I think both of those markets are just tremendously bigger than than I thought three, four years ago. ... I think applying it to inference, I think any inference will work. That's like, you know, there's just so much demand for it. And I don't know if it's gonna get less anytime soon. traffic.libsyn.com
AlvaSupportsHis expectation of sustained, substantial inference demand strengthens the thesis's inference-demand upside.Harrison Chase
Co-Founder and CEO of LangChain
One, I think like the frontier models are relatively getting cheaper over time. ... Two, there's just way more demand. When you start talking about agents talking to agents or multi agent systems, there are just so many more tokens that are getting burned, and so the costs are going up. ... It's actually gone down the last two months, while tokens, the number of tokens that we've used, has actually gone up. traffic.libsyn.com
AlvaSupportsGrowing multi-agent token use supports the thesis's inference-demand runway even as frontier models get cheaper; LangChain's rising token use despite lower coding-tool spend adds a concrete example.
Ben Horowitz
Co-Founder & General Partner of Andreessen Horowitz
And there's a shortage, like a real shortage of GPUs with power and cooling. ... So he he has the advantage of that he's got supply in a great shortage ... We don't really have a lack of demand in AI so I think the dynamics will be different. dts.podtrac.com
AlvaSupportsA shortage of GPUs with power and cooling, alongside his view that AI demand is not lacking, strengthens the demand runway behind the thesis's inference upside.