The age of the GPU is over. The time of the orc…sorry, rack, has come.
Jokes aside, CES 2026 ushered in two major paradigm shifts. Rack-scale solutions are where the industry is headed, and physical AI is likely to consume a large share of the newly deployed capacity. Two neoclouds are already lining up to deploy NVIDIA’s Vera Rubin systems when they land, one of them is looking less like a startup and more like a hyperscaler, and AMD are powering Italian robots.
Because of course they are.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #83
Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.
The Rack Becomes the Product at CES 2026
AMD and NVIDIA both showed the same thing from different angles: GPUs are last year’s news. 2026 is all about the rack.
AMD’s “AI Everywhere for Everyone” leaned into Helios and the MI400 series as a coherent rack strategy. Zen 6 EPYC “Venice” CPUs, up to 72 MI455X GPUs per rack, HBM4 everywhere, and a clear split across MI430X, MI440X, and MI455X based on workload, not SKU sprawl. Low-precision AI at the top end. Mixed AI and HPC lower down. UALink and Ultra Ethernet frame how these systems scale, but the message sits elsewhere. This ships as a system, not a parts list.
NVIDIA came at the same destination from the opposite direction. Vera Rubin NVL72 is now in production, fully rack-native, fully liquid-cooled, and built around codesign across CPU, GPU, DPU, networking, and storage. Fans, cables, and air are gone. Performance gains come from bandwidth, memory locality, and IO offload rather than raw transistor brute force. And Nebius and CoreWeave have already publicly announced deployment plans.
Why this matters:
GPUs stop being the buying decision. Racks, power envelopes, and deployment timelines take their place.
If the vision takes off, infrastructure assumptions get locked in earlier, and new AI data centres win. Legacy sites, on the other hand, face harder retrofits and the risk of stranded assets
AMD and NVIDIA are converging on the same conclusion: define the rack, and you define the platform.
xAI Closes $20B Series E, Locks in Scale
Valor, StepStone, Fidelity, Qatar Investment Authority, MGX, and Baron participated, with strategic checks from NVIDIA and Cisco. The raise backs continued build-out of Colossus I and II, pushes capacity beyond one million H100-equivalent GPUs, and funds Grok’s expansion across consumer, voice, and enterprise surfaces tied to X. Grok 5 trains in parallel with regulators continuing to probe recent content incidents in multiple jurisdictions.
The work continues, but the margin for error tightens as xAI courts enterprise buyers.
Why this matters:
$20bn buys uninterrupted build cycles as xAI moves from consumer to regulated buyers while others wait on power queues, debt markets, or regulators.
Controlling models, product surfaces, and the compute underneath collapses iteration time, potentially assisting in speed of response to enterprise customer demands.
However, this shift comes with risk. Outages, misuse, or compliance slips stop being PR issues and start becoming contractual and legal ones.
Nebius Locks In 80MW Israel Expansion
Nebius has signed a long-term data centre agreement in Israel, securing 80MW across two new facilities built and operated by Mega Or.
The deal covers two sites. 22MW in Masmiyya. 58MW in Beit Shemesh. Total build cost sits at roughly $880m, or about $11m per installed megawatt. Delivery starts in Q3 2026, with phased handover running into Q1 2027. The lease runs for five years, with extension options baked in.
Why this matters:
Nebius is no longer following hyperscalers into regions. It is competing with them, locking in multi-site, long-term capacity in a market where real hyperscalers are already active and competitive.
In doing so, Nebius is buying time and power, not GPUs.
Control of land, megawatts, and delivery schedules are serious growth constraints, so forward planning like this takes some risk off the table.
AI edges closer to regulated front lines
OpenAI and Perplexity are pushing AI deeper into domains where mistakes carry real-world consequences.
OpenAI has launched ChatGPT Health, and is targeting providers and patients with a dedicated environment for health and wellness. Think separate memory, isolated data, and purpose-built encryption. Medical records and wellness apps can be connected, with a view to grounding responses in personal data rather than generic prompts.
Physicians also shaped the workflows, and OpenAI makes it clear, at least in the press release, that conversations are kept out of model training.
Its targeted enterprise platform leans on multi-model routing, cited outputs, and mobile-first access for officers in the field. Bodycam transcripts, scene photos, policies, and incident reports are integrated with no single model dependency. The compliance statements also follow the same playbook as OpenAI: no training on customer data, and twelve months free for qualifying agencies.
Last year, those assurances would have flown.
This year, and in the context of the mega AI labs running out of data, one cannot help but wonder if all these freebies aren’t being paid for in other ways.
Why this matters:
AI is moving upstream into regulated, liability-heavy environments. Where the money is. And the data.
Healthcare and public safety do not tolerate “mostly right.” They demand 100% accuracy, every time. Large language models cannot guarantee that by design.
From these announcements, product adaptation is the solution, along with clear positioning as support tools, not decision-makers. Whether that solution is viable long-term remains to be seen. If not, expect lawsuits.
AMD, GENE.01, & “Body as Compute” for Physical AI
Generative Bionics unveiled GENE.01, its first humanoid robot concept, on stage during AMD’s CES 2026 keynote.
GENE.01, as stated in the press release, treats the body as part of the system, not a shell around it. Full-body tactile skin works as a distributed sensor network. Touch feeds directly into control loops. Perception, motion, and inference run as a single pipeline. Latency matters more than throughput. Near-sensor compute takes priority. AMD supplies the onboard stack across CPUs, GPUs, and FPGA-based embedded platforms, handling real-time perception, sensor fusion, and actuation on the robot itself. Training stays in data centres. Execution moves onto the machine.
Why this matters:
“Body as compute” shifts AI from abstract models to physical systems with the tightest latency and power constraints.
Heterogeneous architectures are moving centre stage. Real-time CPUs, edge GPUs, and FPGAs matter as much as headline FLOPS.
A new compute paradigm is emerging, where intelligence resides inside machines, not just inside racks.
Microsoft Flags a Widening Global AI Divide
Microsoft published its latest Global AI Adoption report, tracking the global spread of generative AI.
Per the report, roughly one in six people worldwide now use generative AI tools. Adoption climbed in H2 2025. Usage across the Global North, however, grew nearly twice as fast as the Global South. By year-end, 24.7% of the working-age population in higher-income economies used AI tools, versus 14.1% elsewhere. Infrastructure and policy also show up clearly in the rankings, and seem to indicate that smaller, digitally coordinated states (UAE, Singapore) are outperforming larger economies. Strong model development and infrastructure presence in the US, for instance, did not translate into widespread adoption, leaving the US outside the top twenty.
Why this matters:
The Global North benefits from dense digital infrastructure, coordinated public-sector rollout, and language-ready products. Much of the Global South faces higher costs, weaker distribution, and fewer locally adapted tools, which slows usage even when models are available.
Free or near-zero-cost tools, open licensing, and offline-friendly deployment matter more in emerging markets than frontier performance, as evidenced by where DeepSeek shows up in the dataset.
Countries with active government adoption, education integration, and national AI programmes convert infrastructure into usage. Markets that leave adoption to enterprises alone lag.
Growth now depends on pairing cloud with local language support, public-sector deployment, and policy alignment. Hyperscalers have the advantage here because they have both the distribution and capital needed to make it work. The question is, do they have the appetite?
Crusoe Gets Green Light for 1.8GW Wyoming Campus
Crusoe has secured approval to build a 1.8GW gas-powered data centre campus in Wyoming.
Per DCD reporting, Project Jade will be developed outside Cheyenne. It pairs hyperscale compute with on-site natural gas generation. Tallgrass plans to invest up to $7bn in up to 2.7GW of generation capacity, supporting an initial 1.8GW IT load. Campus designs target a 10GW long-term ceiling with the first buildings targeting electrification in 2027.
Why this matters:
Gas-first campuses move from workaround to blueprint. Power co-location now leads site selection, not grid queues.
10GW “optionality” is risk management, allowing developers to underwrite future demand without committing full capex upfront.
Wyoming joins Texas as a serious contender for AI megasites, trading decarbonisation optics for speed, certainty, and scale.
The Rundown
“The future is already here, it’s just not evenly distributed.”
That’s how this week felt. Rack scale systems, $20B funding rounds, healthcare and law enforcement AI, and robots on the one hand. A widening adoption gap with the global south on the other.
The technology isn’t the bottleneck anymore.
Power, capital, geography, and governance are.
AI keeps pushing forward, but it’s doing so unevenly, concentrating capability where infrastructure already exists and moving fastest where risk tolerance is highest. If this trend continues, the next phase won’t be defined by which company has the best model, but by which country can deploy, operate, absorb, and risk-manage AI at scale.
That distribution gap is now the story to watch.
See you next week.