The departures come weeks after SpaceX completed its all-stock acquisition of xAI, creating a combined entity valued at $1.25 trillion (Issue #87). Everyone involved is in line for a significant windfall from a potential IPO targeting $1.5T later this year.

It's a natural exit point, for sure. But five departures in twelve months, plus a weekly wave of senior staff at the same time as a major direction change, likely isn't just founders cashing out.

xAI's Grok chatbot has faced regulatory probes in multiple jurisdictions after its image generator allowed mass creation of non-consensual explicit deepfake images. The kind of thing that creates friction on a technical team, emotional and otherwise. That means the product issues and the talent drain may not be unrelated.

We've been tracking Firmus since the Singapore seawater cooling MoU (Issue #51) and the AU$500M Southgate expansion (Issue #74). This is a step change in scale. Now, with $10B in debt, not equity, this is a step change in scale for the Australian neocloud.

This deal allows Firmus to keep its cap table clean while scaling to 1.6GW, approaching the kind of capacity commitments we've seen from CoreWeave and Lambda in the US.

Australia is positioning as a serious AI compute destination. Competitive energy pricing, skilled workforce, and distance from the US-China export control flashpoints make it attractive for customers who need capacity outside traditional hubs. We've tracked the geographic diversification of AI infrastructure across the UK (Issue #86), the Middle East (Issue #74), and the Global South (Issue #44). Australia is the next node.

Submer's evolution mirrors what we're seeing across the infrastructure stack. Companies that started with a single capability, liquid cooling in Submer's case, are acquiring their way into full-stack positions, with each acquisition layer driving down costs on the one hand, and adding recurring revenue and customer lock-in potential on the other.

Radian Arc’s position as a telco embeds GPU infrastructure inside carrier networks. That means data stays in-country, billing runs through telco systems, and latency potentially drops below what centralised data centres can match.

5GW of land and power pipeline is a serious claim. If even a fraction materialises, Submer is operating at a scale that puts it in conversation with the bigger neoclouds. The question, as always, is how much of that pipeline converts to operating capacity.

This is the inference economics thesis playing out in real time. The same issue where Olix raises $220M for photonic inference chips, and Fractile invests £100M in in-memory compute, OpenAI ships a product that enables real-time collaboration - interrupting, redirecting, and iterating at conversational speed. That’s a different product, not just an iterative product improvement.

The $23B valuation in last week’s Series H was built on the thesis that wafer-scale silicon could carve out a meaningful role alongside GPUs. OpenAI's explicit validation of that premise, in production with paying users, is the strongest validation of that thesis yet.

This success feeds into the “two modes” roadmap. OpenAI described a future where Codex blends real-time interaction (Spark-class models on Cerebras) with longer-horizon reasoning (frontier models on GPUs), delegating to sub-agents in the background. If that architecture works, every coding tool will need both a speed tier and a depth tier. The infrastructure requirements and the hardware mix look very different from today's GPU-only stacks.

Both of these startups share a problem focus: memory. Fractile’s solution seeks to tackle the bandwidth issue with extreme compute proximity. Olix's decision to skip HBM entirely goes a step further, and is arguably the most aggressive supply chain bet in the alt-compute field.

With every other chipmaker fighting for the same constrained packaging and memory capacity that NVIDIA and AMD dominate, Olix is betting photonics lets them sidestep the queue altogether. If it works, the supply chain independence will be as valuable as the performance gains.

The inference silicon race keeps intensifying. We covered Mythic's $125M raise for analogue compute-in-memory chips, FuriosaAI landing LG with 2.25x better performance per watt than GPUs (Issue #69), and last week Positron and Cerebras raised $1.23B between them (Issue #87). Now add $320M more for UK-based alternatives. The capital flowing into non-NVIDIA inference hardware has never been higher.

Meta is already showing why this matters: its $60.5B in 2025 earnings included a $2.6B boost from extending estimated useful lives of servers and network assets. One accounting judgment. Material impact on reported profit. New US rules requiring disaggregated expense reporting won't take effect until 2028.

These companies built trillion-dollar businesses on software margins. Now they're spending like railroads but reporting like software companies. The WSJ was explicit in pointing this out: Union Pacific and Norfolk Southern report depreciation as standalone expense lines. The hyperscalers don't.

Useful life assumptions are the hidden lever. If those assumptions prove optimistic, the depreciation catch-up hits earnings all at once.

DataVita is leading the Scottish AI infrastructure charge. On top of the £8.2B AI Growth Zone in North Lanarkshire, and landing CoreWeave as a tenant, DataVita is also winning the bread-and-butter enterprise compute contracts that keep the lights on.

Glasgow specifically cited DataVita's "commitment to renewable, energy-efficient operations" as aligned with the city's ambitions. DV1 runs on 100% green power with a PUE of 1.15. In an era where councils face political pressure over both digital resilience and climate commitments, that combination is a procurement advantage.

The contract is a signal about where Scottish public sector IT is heading. Away from in-house server rooms (Glasgow suffered a data centre outage over a decade ago when a fire suppression system damaged IT equipment) and toward multi-source delivery models with local providers.