There’s a moment in every cycle where the music changes, and everyone in the room feels it before they understand it. That was this week.
Not loud enough for the headlines. Not dramatic enough for the market to price it. But noticeable to anyone who knows what to listen for.
The industry's engine still hums.
But there’s a new texture underneath it.
A hint of slowdown in places that were once pure acceleration. A flicker of momentum where you would not expect it. A quiet reshuffling of who holds the cards of power. And a tell as to who may yet be bluffing.
I’m Ben Baldieri. Every week, I break down what’s moving in GPU compute, AI infrastructure, and the data centres that power it all.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #76
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 US Government’s Genesis Mission Begins
The White House has launched the Genesis Mission, a national program built to stitch every major federal research asset into a single AI-driven discovery engine.
It’s a Manhattan Project-scale attempt to fuse DOE supercomputers, national lab datasets, robotic labs, and secure cloud environments into one closed-loop system for scientific progress.
The Executive Order is explicit.
Unify every DOE high-performance machine. Digitise and standardise the world’s largest scientific dataset collection. Train scientific foundation models across domains like biotech, materials, fusion, quantum, and semiconductors. Run AI-directed experiments inside automated labs.
And do all of it inside a security perimeter built for national defence workloads.
With the order signed, the DOE has:
90 days to map all compute, storage, and network assets across national labs and cloud partners.
120 days to classify and ingest core datasets.
240 days to assess every robotic lab in the system.
270 days to demonstrate an initial operating capability.
That initial capability is designed to pull in academia, defence sites, industrial partners, semiconductor fabs, production plants, and private operators. It is also designed to operate under national security rules, including classification controls, supply chain screening, export restriction compliance, and hardened access for non-federal partners.
Why this matters:
The US is trying to collapse scientific timelines using raw computational force at national scale. Faster drug discovery, new materials, safer nuclear systems, and semiconductor breakthroughs now sit inside one program.
All this sits on top of a year of related moves, including the federal AI Action Plan, directives on youth AI education, orders on export controls, federal AI systems, building national-scale datasets, and childhood cancer research using AI. Genesis is the attempt to bind all of that into one machine.
The nation-state level game has changed. If and how anyone else can respond remains to be seen.
Click here to read the full executive order, and here to read the accompanying fact sheet.
HSBC Sounds Alarm on OpenAI’s Commitments
HSBC updated its outlook for OpenAI this week, and the numbers are brutal.
The ChatGPT maker’s long list of cloud deals with Microsoft and Amazon has pushed its contracted power to 36GW. That means a rental bill heading toward $620b a year once fully online. Only a third of that power arrives before 2030.
To check whether this is survivable, analysts mapped out a decade of revenue growth.
With some very generous assumptions.
They model three billion users by 2030 with ten percent paying. With a two percent share of the digital ad market. And steady enterprise adoption.
All of that still leaves a $207b hole in OpenAI’s finances.
And that’s with a $10b buffer on top.
The conclusion is simple. OpenAI survives only if it keeps raising money at record speed. Or if cloud and chip partners quietly soften the terms of the compute deals the company has already signed. This is unlikely given the scale of their debt commitments, shareholder expectations, and investor concerns regarding wobbles in the private credit market.
With these factors in mind, along with sama’s reaction to simple questions of financial viability and the quiet admittance of “rough vibes” regarding Google’s recent successes in a leaked memo, there are increasingly more questions than answers as to whether OpenAI can pull this off.
Why this matters:
OpenAI’s growth is being propped up by hyperscaler credit, debt facilities, and the hope that its user base converts fast enough to keep the machine running.
A liquidity squeeze would force renegotiations with Microsoft, Amazon, and anyone else supplying compute. In this scenario, neoclouds are the least likely to offer flexibility given their concurrent debt loads, comparatively thin margins, and inability to subsidise revenues from other business units.
OpenAI’s success is tied to the success of all the major players in the industry. If a black swan hits the market, stability would not be guaranteed. And that would be bad for everyone, everywhere, all at once.
Megaport Officially Closes Latitude.sh Acquisition
Megaport has officially closed its acquisition of Latitude.sh, as confirmed by CEO Michael Reid.
Latitude brings the muscle. Instant GPU and CPU deployment. Automated bare metal. Rapid provisioning. A clean developer API. High-density racks already serving AI inference and training workloads.
They built a global niche by making compute feel like a button press.
A private network spanning more than a thousand data centres. Direct routes into AWS, Azure, Google Cloud, Oracle, Alibaba, IBM, and every major SaaS endpoint. Low-latency interconnect. Elastic bandwidth. Enterprise-grade contracts. A footprint few operators can match.
Together, they become something new:
A network-native compute cloud that can deliver compute anywhere that Megaport already has fibre.
Why this matters:
This acquisition is a big deal (pun intended) for both companies. Latitude gains global distribution overnight, solving the hardest part of its business: reach. At the same time, Megaport gains an actual compute layer, letting customers deploy AI workloads directly into its backbone rather than routing traffic elsewhere.
The subsequent combined platform includes GPUs, CPUs, interconnect, and multi-cloud routing, all available in one place, creating at least a partially credible alternative to hyperscalers for latency-sensitive inference and training.
As such, the midmarket neocloud space just got a lot more competitive. Expect further consolidation in the near future.
Anthropic Releases Claude Opus 4.5
Anthropic released Claude Opus 4.5 and aimed it straight at last week’s Gemini 3 rollout.
Opus 4.5 leans into real work. Agent runs last longer. Tooling is sharper. Memory stops collapsing on long jobs or chats, and effort controls let teams pick speed or depth. Claude Code also produces clearer plans, and Chrome, Excel, and desktop support expand.
And, on top of all that, it’s cheaper while staying competitive.
Pricing drops to 5 dollars in and 25 dollars out per million tokens with solid benchmark performance.
Why this matters:
Gemini still leads the charts, but Opus makes the economics work for teams running constant, long-lived agent jobs.
As more workloads shift to chained tasks and higher utilisation, stability and cost discipline matter more than peak benchmarks. Until intelligence converges with the cost of energy, that is. Then it’s probably back to benchmarks.
We’re in another round of “our best model yet”, so expect some creative PR from the competition in the coming weeks.
Meta in Talks with Google for TPU Deployments
The pressure kicked up after reports that Meta is looking at deploying Google’s TPUs from 2027. Ironwood, Google’s seventh-generation TPU, is a serious platform. It scales from 256 to more than 9,000 chips in a single pod with optical circuit switching. It matches the scale of top GPU clusters and, in some cases, beats them. And the biggest shift of all is Google’s willingness to place TPUs directly inside customer data centres.
No longer cloud-only.
Now on-prem.
That is a direct challenge to NVIDIA’s core business. When the news broke, Team Green jumped on X to insist it still delivers stronger performance, broader compatibility, and wider model support. That reaction tells you everything you need to know.
Google is no longer nibbling at the edges.
And others are already moving.
Anthropic plans to use up to a million TPUs for future Claude systems, matching its Trainium deployments. FluidStack is building data centres for Anthropic and deploying capacity into Terawulf and Cipher Mining’s data centres with Google’s financial backing.
To be clear, NVIDIA still dominates.
But Google is building a real counterweight with scale, hardware, customers, and distribution.
Why this matters:
Google is pushing its alternative accelerators into customer facilities, not just its own cloud, putting direct pressure on NVIDIA’s strongest market: hyperscalers and hyperscale-like players.
Interest like this from first Anthropic, and now Meta, coupled with the launch of Gemini 3 (which was allegedly trained entirely on TPUs) is the strongest indication we’ve had that GPUs are no longer the only game in town.
To that end, homogeneous silicon and single-vendor strategies could become increasingly risky as hyperscalers and neoclouds spread their bets across TPUs, GPUs, and ASICs. Change is afoot.
Google Targets 1000x More Compute in 5 Years
Moore Google news, as the company also told its employees that it now has to double AI serving capacity every six months just to keep up.
Amin Vahdat, who runs AI infrastructure at Google, laid it out at an internal meeting:
Demand curves are steepening, supply is the bottleneck, and the company needs a thousand-fold jump in capability over the next four to five years without blowing out power budgets.
Google’s response to this crunch leans on three levers:
More efficient models (like Gemini 3 last week)
Aggressive infrastructure build-outs (still to be seen)
CEO Sundar Pichai acknowledged market anxiety around a potential AI pullback but stressed the bigger risk is underinvesting. He pointed at cloud revenue climbing 34 percent year-on-year and a 155 billion-dollar backlog as evidence the demand curve remains intact.
Why this matters:
The pace required to meet demand is now beyond linear planning.
If the six-month doubling holds, supply-chain strain, power scarcity, and land constraints move from background noise to existential.
Google is effectively saying the floor for hyperscaler buildouts is rising again, and the next wave will be dominated by custom silicon and energy efficiency, not just GPU volume.
AWS to Build ~1.3GW of Capacity for US Government
AWS just placed one of the biggest federal AI bets of the year.
The company will invest up to $50b to build dedicated AI and supercomputing infrastructure for US government agencies. The buildout will add nearly 1.3GW of classified and unclassified compute across AWS Top Secret, Secret, and GovCloud regions. This includes new data centres, new networking systems, and full-stack access to SageMaker, Bedrock, Nova, Anthropic Claude, open-weighted models, and NVIDIA systems.
AWS frames this as the next phase of government cloud, shifting from isolated HPC sites to something far larger:
An AI-integrated national grid where modelling, simulation and real-time feedback loops run inside secure regions tailored to different classification levels, purpose-built for missions that cannot fail.
HPC for defence. Long-context modelling for intelligence. Supply-chain optimisation for federal agencies. AI steered experimentation for energy, healthcare and materials. Autonomous analysis of satellite imagery and threat patterns.
All with timing aligned to the Administration’s AI Action Plan and the new Genesis Mission, giving the federal stack a direct commercial partner with the experience needed to deliver at the scale required by a vision of unquestionable US AI dominance.
Why this matters:
1.3GW of new capacity sets an aggressive benchmark for classified AI infrastructure and tightens the link between federal demand and hyperscaler design choices.
AWS is locking in long-term control of the federal AI footprint at the moment the US government is standing up its largest scientific compute program in decades.
Hyperscalers, sovereign clouds and private platforms are now converging on the same workload class: mission-critical AI running at national scale, where reliability, security, and national pride matter more than price.
The Rundown
A second big week for Google shows the search giant is gaining ground on two fronts at once.
On one side, NVIDIA. A major customer is exploring Google’s silicon for future clusters. TPU pods are scaling past 9,000 chips. The hardware is now landing directly inside customer data centres, not just behind Google Cloud’s walls.
The response from Team Green carried the tone of a company that knows the threat is real.
On the other side, OpenAI.
Financial strain is now being quantified in public. Their own leaked internal memo admits uncomfortable “rough vibes”. Meanwhile, Google continues stacking wins: a model release that beat expectations, TPU adoption from Anthropic, on-prem interest from competitors, and infrastructure leadership openly discussing doubling serving capacity every six months.
Now, they’re taking on the two biggest names in the industry at the same time.
None of this says the race is over. Far from it.
But, it does suggest something uncomfortable for two of the industry’s biggest names:
Google’s steady, unglamorous, deeply integrated approach is working.
And the longer the pressure builds, the harder it becomes to argue otherwise.
See you next week.
Everything Else
No everything else this week as I’ve been hit with flu number one of the year.
Normal programming will resume next week!