We may look back on this week as the week that Google won.
Not because they shouted about it. Not because they blasted out PR on every available channel. But because they beat out the competition to underpin Apple’s new Foundation Models across every single iPhone.
Distribution is king.
Always has been. Always will be.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #84
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.
Google’s Gemini to Power Apple Intelligence
Sometimes, the biggest moves are the quietest.
Apple is building its next generation of Apple Foundation Models on top of Google’s Gemini models and infrastructure. Per the Google announcement, these models will power future Apple Intelligence features, including a more personalised Siri, while continuing to run on-device and via Apple’s Private Cloud Compute. Apple evaluated multiple providers before settling on Google, citing model capability rather than cost or convenience. Details are sparse, but even so, selection is a massive vote of confidence in the search giant’s AI capabilities, and a serious issue for competitors.
Why this matters:
Apple doesn’t need to win model benchmarks. It needs reliable intelligence across a billion devices. Gemini now sits under that layer.
In turn, Google gains what no AI lab can buy: default distribution at iPhone scale, embedded into daily behaviour rather than optional usage.
Couple this with further distribution on Android, and that’s both pillars of the market, meaning the mobile AI race might be over before it’s really even begun.
Trump Slaps 25% Tariff on H200s & MI325Xs
The semiconductor tariff game continues, and now we know how payment will actually work.
The White House has operationalised its 25% tariff on NVIDIA and AMD chip sales to China. Per reporting from the FT, the mechanism applies when chips are imported into the US and then transshipped abroad, effectively converting export control relief into a taxable event. US companies seeking to export GPUs are also subject to the new rules, whereas chips destined for US data centres and the domestic buildout remain exempt. Naturally.
Why this matters:
With export controls now a monetisation tool, access to frontier chips becomes something the Federal Government prices, not just permits or denies.
This sets an interesting precedent. If tolerated, future AI trade may involve revenue-sharing with nation states, not just compliance with rules.
What this means for smaller exporters, the second-hand market, and overall market dynamics remains an open question.
Polarise Secures €117M to Build Europe’s AI Factories
A new neocloud is taking shape in Germany.
Polarise has secured up to €117 million in financing from Macquarie to fund AI data centre fit-outs in Munich and expand GPU infrastructure across Europe. The capital backs Polarise’s AI Factory model, with a focus on sovereign compute and the same dense GPU deployments that the industry is known for. Early 2026 is the target for delivery, with Macquarie structuring the financing through its specialised asset finance arm, prioritising speed and execution to meet the tight timeline.
Why this matters:
There’s a gap in the European market following on from Rumble’s acquisition of Northern Data. Polarise could step up to fill that gap.
Germany’s AI demand already exists (just look at the NVIDIA Telekom partnership). Automotive, industrial automation, defence, and healthcare workloads need sovereign capacity to meet customer and regulatory requirements.
If Polarise is successful, they’re in a strong position to satisfy that demand at exactly the moment that Europe really seems to be waking up to the importance of sovereign capability.
Basecamp Research Brings AI to Genetic Engineering
The EDEN models generate large DNA insertions at precise genomic sites without cutting DNA, moving beyond CRISPR’s small-edit limits. Trained on a massive proprietary evolutionary dataset and scaled on NVIDIA infrastructure, the same system also designed antimicrobial peptides with a 97% lab-confirmed hit rate against drug-resistant bacteria. Lab results show successful insertions across thousands of disease-relevant loci, including CAR-T constructs with strong tumour-cell kill rates. NVIDIA’s venture arm also joined as an investor following deep technical collaboration.
Why this matters:
Basecamp Research trained EDEN on one of the largest proprietary evolutionary DNA datasets assembled, spanning extreme environments and non-model organisms. That dataset encodes billions of years of functional solutions. Models trained on it stop guessing and start exploiting biological priors directly.
Precise, large insertions without double-strand breaks remove a core constraint that shaped CRISPR-era biology. If insertions scale reliably, gene therapy design shifts upstream into software, shrinking iteration cycles, manufacturing complexity, and downstream safety risk.
As Basecamp unlocks more of its dataset with AI, biology starts to resemble a compounding data asset. Each model improves the next, and the limiting factor becomes compute, not wet-lab capacity.
OpenAI Signs 750MW Cerebras Deal
Alternative semiconductors are having a moment
OpenAI has signed a multibillion-dollar capacity deal with Cerebras Systems, committing up to 750MW of compute over three years. The agreement, valued at north of $10b, will see the systems deployed to focus real-time inference workloads where response speed directly drives usage and revenue. Cerebras’ wafer-scale architecture, long positioned as exotic, now moves into frontline deployment.
Why this matters:
Following on from NVIDIA’s Groq acquihire, this deal sees Cerebras join a short list of inference-first players gaining serious buyer attention.
This makes sense as performance, latency, long GPU lead times, and grid constraints are increasingly forcing buyers to diversify away from homogeneous deployments.
What’s not clear from this announcement, however, is when it will come online. “Through 2028” is ambiguous at best, and given how fast this market moves, a lot can happen in 3 years. By the time this capacity is fully online, the competitive landscape could look very different.
Meta Compute Puts Power and Capacity Back in Zuckerberg’s Hands
Mark Zuckerberg says Meta plans to deploy tens of gigawatts this decade, with ambitions that could stretch into the hundreds over time. The effort consolidates data centres, silicon strategy, energy procurement, and supplier partnerships under one roof. This follows a year where Meta committed roughly $72b in capex and locked in long-term nuclear power deals to keep future clusters fed.
Why this matters:
Owning power strategy, capacity planning, and infrastructure execution shortens feedback loops and reduces exposure to third-party bottlenecks.
After Llama’s stumbles, Meta is betting that infrastructure control, not just model tuning, becomes the differentiator. In this scenario, compute abundance becomes the hedge against catching up late.
“Hundreds of gigawatts” sits closer to national grid planning than cloud expansion and forces the conversation with utilities, regulators, and suppliers. Assuming Meta can both fund these ambitions and secure buy-in.
Soluna, HydraHost Start Talking Hyperscale Numbers
More neoclouds are beginning to frame capacity in hundreds of megawatts rather than racks or pods.
Soluna and Metrobloks plan a 100+ MW AI and HPC build at Project Kati 2 in South Texas, with an expansion roadmap past 300 MW. In parallel, Bitzero is bringing a 200 MW Finland site to market via CBRE. Power is scheduled for Q1 2027, and Hydra Host is lined up to monetise early GPU capacity across hyperscale and neocloud demand channels.
Both efforts target the same gap: GPU-hungry customers stuck between hyperscaler queues and sub-scale regional builds.
Why this matters:
100-300 MW announcements from groups like Soluna, Metrobloks, and Bitzero would have been unthinkable for anyone outside the hyperscalers 18 months ago.
However, none of this is fully locked yet. These are MOUs, brokerage mandates, and non-binding LOIs, not permits, PPAs, or signed leases.
Even so, the fact that these numbers are being discussed at all hints at an execution-dependent market recalibration.
The Rundown
And just like that, the mobile AI race could be over.
Apple quietly chose Google this week. And Google announced this via a two-paragraph post on their corporate website. It’s refreshing to see an actual industry-shifting announcement without all the bluster and fanfare we’ve become so accustomed to. Especially one that would have been seen as completely impossible just a couple of years ago, with the disastrous launch of Bard and subsequent pushback.
Oh, to have been a fly on the wall inside OpenAI when this news broke.
See you next week.