The Signal — July 27, 2026
Two of the biggest names in AI silicon spent the last few days making the same argument from opposite ends of the stack, and a third company made a smaller move worth watching. Here's what moved.
AMD keeps chipping at the CUDA moat
AMD wrapped its Advancing AI 2026 event over the weekend, and the pitch was the one it has been building toward for two years: a full-stack alternative to NVIDIA that customers can actually deploy. Lisa Su put the Helios rack-scale system at the center, alongside the Instinct MI450-series ramp, the new EPYC "Venice" server chips on Zen 6, Pensando networking, and a ROCm software stack that AMD needs the developer world to take seriously. Su told CNBC the company is seeing roughly 30x more performance with Helios compared to its previous generation.
The reason any of this matters is the software lock-in, not the hardware. NVIDIA's real advantage has never been just faster GPUs; it's CUDA, the programming layer that a decade of AI code is written against. AMD can match NVIDIA on paper and still lose if developers have to rewrite everything to switch. So the ROCm push, and the recently announced Anthropic partnership to deploy up to 2 gigawatts of MI450 GPUs, are AMD buying its way into the ecosystem the hard way, one large customer at a time.
The keynote reads differently against sharper coverage. SemiAnalysis, which follows this market closely, describes a messier reality behind the launch: unstable internal development clusters and a difficult MI455X production ramp, propped up by aggressive discounting to move product. None of that means the strategy is failing. It means the gap between a strong keynote and a shipping, developer-trusted platform is still real, and AMD is spending heavily to close it.
Sources: AMD · SemiAnalysis · ServeTheHome
NVIDIA is using its own CPU to design its next chips
While AMD works on breaking into NVIDIA's territory, NVIDIA is quietly expanding into AMD's. The company said it is deploying its next-generation Vera CPU internally to speed up electronic design automation, the software engineers use to design chips, working with the two dominant EDA vendors, Cadence and Synopsys. NVIDIA reports up to 1.5x speedups on selected Cadence Jasper and Synopsys VCS workloads running on Vera.
There is a neat recursion here: NVIDIA is using AI-era silicon to design the next generation of AI-era silicon, and doing it on a CPU it built itself rather than an Intel or AMD part. That last detail is the strategic one. NVIDIA has historically owned the GPU and left the CPU layer to AMD, Intel, and Arm-based designs. Vera, which CNBC detailed on July 21 as part of the broader Vera Rubin platform, is NVIDIA moving into the server-CPU business its rivals depend on.
The performance numbers deserve the usual caution. These are vendor-selected tests on workloads NVIDIA chose, not independent benchmarks, so treat the 1.5x as a directional claim rather than a settled result. The more durable takeaway is what the two chip stories say together: the competitive line between GPU makers and CPU makers is dissolving, and both companies are now trying to own the whole system.
Midjourney bought an astrology app
Midjourney, best known for its image generator, acquired Co-Star, the personalized-astrology app, in a deal reported by Bloomberg last week and confirmed by TechCrunch and The Verge. Terms were not disclosed, and the astrology part is mostly a distraction. What matters is the shape of the move.
Co-Star comes with something a model lab usually has to build from scratch: a consumer product with an established, engaged user base and a daily habit attached to it. Buying it reads as Midjourney deciding its next phase of growth is about owning consumer-facing apps and distribution, not just licensing a very good image model to other people's products. Frontier labs keep discovering that the model is the easy part and the surface area where users actually live is the hard part. This is a small, concrete bet on that idea.
Sources: Bloomberg · TechCrunch · The Verge
On the Editor's Desk
A few things stayed out. Sakana's Fugu-Cyber benchmark story and a UK and US evaluation of Kimi K3's cyber abilities both landed in the same benchmark-caveat lane we covered late last week, so running them again would have been repetitive. The NYT piece on Silicon Valley splitting over Chinese open models is good analysis, but the open-model and China policy thread ran here on Friday and there was no new development to add. A batch of research papers on agent training and world models were solid but too narrow for a general reader today.
