Happy New Year from The GPU! I hope you had a chance to catch your breath over the holiday season, because 2026’s AI infrastructure sprint has already begun!
We’ve got Meta behaving like the Facebook of old. We’ve got Masa taking globally significant digital infrastructure investors private. We’ve got private companies deploying at a multi-gigawatt scale. And we’ve got a neocloud spinning out the compute business unit into a new neocloud via an RTO. And more.
I was concerned this would be a slow news week.
Clearly, those concerns were misplaced.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #
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.
SoftBank Buys DigitalBridge for $4bn
SoftBank Group is acquiring DigitalBridge for $4bn, taking one of the world’s largest digital infrastructure investors private.
DigitalBridge manages $108bn across data centres, fibre, towers, edge, and connectivity assets. The deal gives SoftBank direct control over the physical layer behind AI: land, power, fibre, and financing. DigitalBridge will continue to operate independently under CEO Marc Ganzi, with SoftBank supplying capital, balance sheet strength, and long-term mandate.
Why this matters:
SoftBank is shifting from funding AI models to owning the ground they run on. Compute scale now starts with asset control, not capex access.
Hyperscalers are buying upstream. Google bought Intersect. SoftBank bought the allocator. The power scramble is moving into M&A.
This aligns cleanly with Masayoshi Son’s ASI thesis: whoever controls capital, power, and deployment velocity controls the AI platform layer.
Applied Digital Spins Out Cloud Arm into ChronoScale
Applied Digital is spinning out its cloud business and combining it with Ekso Bionics to create a new Nasdaq-listed GPU platform called ChronoScale.
Applied Digital keeps its data centre ownership and development business. ChronoScale houses the accelerated compute platform. After closing, Applied Digital expects to own about 97% of the combined entity, with completion targeted for the first half of 2026. The focus stays on dense AI training and inference cluster deployments, while leveraging direct access to Applied Digital’s AI factory campuses.
Why this matters:
GPU cloud platforms and data centre real estate are being separated to allow different capital structures, growth paths, and risk profiles.
If this is anything more than an RTO with a faster path to public markets, Ekso’s background in robotics and human-machine systems could indicate a positioning for physical AI.
Similar signals are emerging elsewhere. Brookfield partnered with Figure. Nebius recently ran awards specific to physical AI and robotics. In concert, these suggest a growing appetite for compute tied to embodied systems rather than pure software agents. Though Ekso’s 3% stake could say otherwise.
Meta Acquires Chinese Agent Startup Manus
Meta is acquiring Manus, the Chinese-founded AI agent startup, in a deal Reuters sources peg at $2-3bn.
Manus went viral earlier this year after claiming a general AI agent capable of planning and executing tasks with minimal prompting. The company then moved its HQ from China to Singapore, lowering geopolitical friction and making a US acquisition viable. Meta will operate and commercialise Manus directly, folding it into Meta AI and its consumer and business products.
Why this matters:
This mirrors Meta’s old Facebook playbook: buy emerging competitors early, integrate fast, and remove future threats before they mature.
This follows on from NVIDIA’s Groq acqui-hire, indicating that the big players are increasingly choosing quiet consolidation over open competition.
If the trend continues, the AI market could be entering its defensive acquisition phase, where incumbents absorb talent and capabilities while doing their best to either avoid regulatory scrutiny or rely on political relationships that mean the scrutiny no longer matters.
xAI Targets Enterprise Customers with New Releases
xAI has launched Grok Business and Grok Enterprise, pushing Grok from a consumer tool into the workplace.
Grok Business targets small and mid-sized teams with shared workspaces, higher rate limits, consolidated billing, and admin controls. Grok Enterprise adds SSO, directory sync, and audit controls. An optional Enterprise Vault also includes a dedicated data plane and customer-managed encryption keys.
The pitch is familiar but explicit:
No training on customer data, permission-aware access to internal tools like Google Drive, citation-backed answers, and support for agent-style workflows over large document sets via Projects and Collections.
Why this matters:
xAI is moving up the stack, from model and API to seats, budgets, and procurement. This puts Grok in direct competition with ChatGPT Enterprise, Copilot, and Claude for enterprise workflows.
Enterprise Vault signals where buyer anxiety still sits. Isolation, keys, and data planes now decide deals as much as model quality.
This also pairs cleanly with xAI’s infrastructure buildout. Owning models, software, and increasingly the compute underneath shortens the loop between product demand and capacity planning.
China’s Lisuan G100 GPUs Start Shipping
China finally has a GPU with real domestic momentum.
Per Tom’s Hardware, Lisuan’s G100 series has begun shipping to customers, marking the closest the country has come to a commercial NVIDIA/AMD rival. The first batches are heading to “digital twin” and industrial clients, with early benchmarks suggesting performance in the NVIDIA/AMD mid-tier range. Retail availability is expected in Q1 2026, following mass production that kicked off in September.
Lisuan also claims one noteworthy edge:
Neither NVIDIA nor AMD has that today. In China, where ARM PCs are proliferating, that matters.
Why this matters:
China is edging closer to a credible mid-tier GPU ecosystem that doesn’t rely on NVIDIA or AMD.
ARM GPU support gives Lisuan a differentiated foothold in a fast-growing domestic segment.
If the retail versions land anywhere near the current performance claims, Lisuan becomes the first Chinese vendor to offer a commercially viable alternative to NVIDIA’s 60-series class.
Microsoft CEO Sees AI as New “Bicycle for the Mind”
Satya Nadella has started blogging again and keeps returning to one line: “bicycles for the mind.”
It’s a deliberate callback to Steve Jobs framing of computers as force multipliers for human effort. Not replacements, but tools that let people go further with the same inputs, framing AI as an assistant rather than a substitute.
There’s also a clear avoidance.
No sparring over “AI slop.”
The Verge picked up on this, framing the post as a step around the slop-versus-sophistication debate rather than into it. Nadella doesn’t defend low-quality output. He doesn’t attack it either. He treats it as a symptom of misapplied capability, then moves on.
Why this matters:
If we take this blog post as a statement of intent, Microsoft is shifting focus from model competition to system ownership.
“Bicycles for the mind”, therefore, puts humans back in the loop by design, meaning tighter guardrails, clearer intent capture, and less tolerance for free-form generation.
If reports, however, that no one is actually using CoPilot are true, there may be more urgency to this blog post than is written.
xAI Pushes Memphis Campus Toward 2GW
xAI has bought a third building to expand its Memphis footprint, pushing planned training capacity toward ~2GW.
Per Reuters, the site sits near the Colossus cluster, with conversion into a data centre expected to start in 2026. The build lines up with plans to scale toward a million GPUs, backed by on-site power, including a nearby natural gas plant under development. Musk confirmed the purchase on X, naming the building “MACROHARDRR”.
Why this matters:
This is hyperscale infrastructure without a hyperscaler balance sheet. Private labs are now building at utility-grade scale.
Power-first design is becoming the gating factor. Proximity to generation matters more than cheap land.
The arms race is shifting from models to sites. Control of compute, power, and timelines now defines competitive edge more than model releases.
The Rundown
And so it begins.
If the first issue of The GPU in 2026 is anything to go by, it seems the ever-increasing rate of acceleration is set to continue. Though perhaps with a different tone.
Two of the biggest stories over the Christmas period were acquisitions.
And they weren’t small.
NVIDIA licensed Groq’s tech and hired top execs for $20B. Meta bought Manus for $2-3B. The acquirers are systemically important to the industry. The acquirees are relative upstarts that, given enough time, could have become relative thorns in the side of the monoliths.
That’s likely why they became targets.
And it’s also why we’ll likely see similar deals continue over the next 12 months.
We’re moving out of the “throw things at the wall and see what sticks/fund everything” phase of the market and into a period where the survivors look like targets. Those teams have taken the early-stage execution risk and will likely be rewarded if whatever they’ve built is a viable solution to a problem. Or, if their solution could become a problem for someone else. Unless the correct action is taken, of course.
Whatever happens, it’s going to be a hell of a year.
And I’ll be right here with you every step of the way.
See you next week!