The irony writes itself.

For decades, Washington slammed Beijing for subsidies, protectionism, and hand-picked national champions.

Now?

The US is playing the same hand.

Intel looks more like Huawei than Hewlett-Packard, propped up by state equity, defence contracts, and tariff shields.

Meanwhile, Beijing, once the fortress of joint-ventures and firewalls, now runs an AI market that looks more like Silicon Valley in its prime: cutthroat competition, open-source experimentation, and engineers chasing upside rather than subsidies.

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.

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.

Washington spent decades railing against Beijing’s protectionism. Now it’s playing the same hand.

Intel just secured $2B from SoftBank, and could soon see the US government itself step in as a shareholder. How? Converting CHIPS Act grants to equity. If the deal goes through, the federal government becomes the largest shareholder. Throw in defence contracts keeping fabs alive, tariffs cutting out rivals, and a new red scare around Chinese open-source models developed at a fraction of the cost of domestic equivalents, and you have the makings of an industrial strategy the US once denounced.

Unfettered free market competition is out; tariffs, subsidies, government ownership, and strategic consolidation around national champions are in.

It’s a mirror image of what Beijing did in the 2000s and 2010s.

China forced joint ventures, capped foreign ownership in critical industries, and mandated domestic procurement for state-owned enterprises. Entire sectors were effectively off-limits to overseas suppliers until a local ecosystem could stand on its own. Huawei grew under state contracts. Baidu and Tencent scaled behind the firewall that kept out Google and Facebook. SMIC was shielded from competition while absorbing foreign IP and talent.

It was messy, uneven, and state-heavy, but it worked.

The irony is that Beijing, to a degree, has since moved on.

China’s AI sector today looks more like Silicon Valley in its prime: cutthroat competition, rapid iteration, and a near-monopoly on untapped pools of elite engineering talent. Coupled with an overbuilt power grid that needs more demand, and an open-source AI ecosystem aligned to national priorities, the state has created a protected market that breeds a level of ferocious internal competition Reagan would be proud of.

Why this matters:

Read the full Intel/SoftBank story on Reuters here.

Crusoe has acquired Atero, seeking inference efficiency.

Atero’s tech optimises GPU utilisation and memory placement for LLMs and other demanding AI tasks. By integrating it directly into Crusoe Cloud, the company aims to improve performance, reliability, and energy efficiency across its managed AI services. The Tel Aviv base also gives Crusoe access to Israel’s talent pool and regional customers.

Why this matters:

Click here to read the announcement from Crusoe.

Fluidstack has wasted no time expanding its foothold in New York.

Just a week after inking a $3.7B, 200MW hosting deal at TeraWulf’s Lake Mariner campus, the company has exercised its option to add another 160MW. The move brings its total contracted capacity at the site to 360MW, translating to $6.7B in guaranteed revenue, and potentially $16B if extensions are triggered. To support the expansion, Google is raising its financial backstop to $3.2B and upping its stake in TeraWulf to 14%, further cementing its role as the anchor partner behind Fluidstack’s growth.

Why this matters:

Click here to read the full story on DCD.

IBM and NASA just open-sourced Surya, the first foundation model for heliophysics.

It ingests 9 years of NASA Solar Dynamics Observatory data to predict solar storms, flares, and coronal mass ejections that can knock satellites out of orbit, disrupt GPS, and strain power grids. The model is now live on Hugging Face, GitHub, and IBM’s TerraTorch, alongside SuryaBench, a curated dataset and benchmark suite designed to standardise AI-driven solar research.

Why this matters:

Click here to read the full announcement from IBM.

SoftBank has acquired Foxconn’s 6.2m sq ft Lordstown facility for $375m.

The tech group is planning to convert the Ohio plant from EV production to AI server manufacturing for the $500bn Stargate project. Foxconn will continue to operate the plant, retrofitting it to produce servers for SoftBank, Oracle, and OpenAI’s planned US megadata centres. The Lordstown site's massive power capacity and room to expand make it a natural anchor for Stargate’s hardware pipeline. It’s also one of the largest facilities of its kind in the US; six times bigger than Foxconn’s Houston AI server site.

Why this matters:

Click here to read the full story from Tom’s Hardware.

At this week’s Made by Google 2025, the headline wasn’t the hardware.

It was the distribution.

Google is putting AI straight into the channel that matters most: Android phones already in users’ hands. The new Pixel 10’s Tensor G5 chip is the first designed to run Gemini Nano natively. That means translation during calls, automatic summaries, live camera guidance, and local text/image generation, all happening on-device. Pixel Pro buyers get more: bundled access to AI Pro with Imagen 4, Veo 3, and other generative tools.

On top of that, AI Mode in Search has gone global, with new agentic features now live in 180 countries.

Together, it’s a dual-layer strategy:

Why this matters:

Click here to read the AI mode update from Google.

Lambda is adding 30MW+ of dedicated data centre capacity.

EdgeConneX will build two new AI data centres for the neocloud, in Chicago (23MW) and Atlanta. The facilities are designed for 600kW+ racks with hybrid liquid-to-the-chip and air cooling, positioning them for the next wave of GPU-dense clusters. The builds expand Lambda’s footprint beyond its leased colo sites and push it closer to its stated 2GW, 1M GPU target by 2030. For EdgeConneX, the deal reinforces its “build-to-density” strategy, delivering single-tenant, AI-optimised infrastructure at speed for hyperscale-style buyers.

Why this matters:

Click here to read the full press release from Lambda and EdgeConnex.

Washington and Beijing may be swapping costumes, but the stage is the same.

The US is still the innovation powerhouse. Google embedding Gemini Nano on Tensor G5 into every Pixel is a distribution play nobody else can match. Fluidstack at Lake Mariner, Lambda with EdgeConneX; the sheer pace and scale of data centre buildouts in the US remains astonishing. And you can’t ignore the software gravity pulling everything into Hugging Face, or the commercial firepower behind projects like Stargate.

But it’s interesting to watch how industrial policies, East and West, are coalescing into a similar model: a pyramid with the same shape, but different apexes.

In Beijing, capital serves society. The state directs money, talent, and infrastructure into national projects, with individuals positioned to support the system.

In Washington, society serves capital. The state (apparently) props up firms, but the flow of value is designed to reward investors, shareholders, and markets first.

The open question is what role the individual has inside either system - worker, beneficiary, or afterthought?

Because the real contest isn’t just whether AI scales.

It’s who the pyramid is ultimately built for.

And what that means for the rest of us.

See you next week.