Two Australian neoclouds making major capital markets moves in back-to-back weeks. Firmus locked in $10B in project debt last issue (Issue #89). Sharon AI has taken the public equity route. With Firmus deploying thousands of GPUs across a 1.6GW national rollout, and Sharon AI going public on 432 GPUs with near-total customer concentration, the contrast is stark.

The prospectus risk factors read like a checklist of neocloud fragility: limited operating history, ongoing losses, capex intensity, supply chain exposure, and significant customer concentration. And while extreme customer concentration isn’t necessarily a deal-breaker in this industry (just look at NVIDIA), losing one would likely have a material impact. Sharon AI openly acknowledged this.

Public markets unlock deeper liquidity and lower capital costs on the one hand, but quarterly reporting forces transparency on metrics that private competitors never have to disclose on the other. That's the tension in taking a neocloud public. So, credit where credit is due to Sharon AI for choosing this path. If they succeed, we may yet see the floodgates open for small-neo IPOs.

Mistral's trajectory tells a story about where the value is moving. The company started as a pure model builder. Then it was a FluidStack customer renting GPU capacity (Issue #23). Then it became an NVIDIA Cloud Partner. Now it's acquiring infrastructure teams, investing in its own data centres, and working with HSBC on enterprise deployments (Issue #77).

Koyeb's serverless platform solves a specific problem Mistral couldn't build fast enough internally: on-premises deployment and GPU orchestration for enterprise clients. As Mistral pushes deeper into regulated industries, the ability to deploy models on a customer's own hardware becomes the dealbreaker.

The political tailwind behind non-US AI infrastructure is starting to generate real M&A activity, not just press releases.

NVIDIA has long been investing in customers, backstopping loans, and creating circular financing structures to lock in GPU orders. AMD doing the same thing confirms that this route to market is becoming the go-to model for chip vendors competing for neocloud deployments. The chip sale is now a financial product, not just a hardware transaction.

The economics tell you how hard AMD is pushing. A 6% rate on GPU-backed debt is cheap. If the reporting is true, AMD effectively had to put its own balance sheet on the line, offering to lease back chips it already sold, to get Goldman comfortable at that price. That's a vendor saying, "We believe in this hardware enough to eat it if nobody else wants it."

Bold. Especially when it also means AMD is carrying contingent liability for Crusoe's commercial success.

Two major model upgrades in one week, each deployed to hundreds of millions of users across multiple product surfaces. Every generation that ships drives another wave of inference demand, and both companies are compressing the gap between releases to months, not quarters.

Price-performance keeps collapsing. What required Anthropic's Opus-class pricing three months ago now ships at Sonnet rates. For API customers buying inference at scale, the cost per unit of useful intelligence is falling fast. That's deflationary pressure on inference revenue across every provider hosting these models.

OpenAI took a different approach entirely this week. Rather than shipping a new model, it hired OpenClaw creator Peter Steinberger to build consumer-facing agents. Because sometimes, hype and ready-made distribution are more important than rankings.

An 8 exaflop deployment in India is the second major Cerebras validation in two weeks, after OpenAI shipped its 1,000 tk/s Codex-Spark model on Cerebras hardware (Issue #89).

India is suddenly at the centre of multiple AI infrastructure plays in the same week. Submer's Radian Arc signed two Indian partnerships to embed GPU edge infrastructure in telecom and colocation networks (see below). Now G42 is deploying exaflop-scale national compute. The common thread is sovereignty: every deal is structured around Indian data residency, Indian governance, and Indian access.

The UAE-India corridor is becoming a significant AI infrastructure axis. Abu Dhabi is positioning itself as a sovereign AI infrastructure exporter, building compute capacity in partner nations rather than asking those nations to send their data to the Gulf. That's a different model from the US hyperscalers and, for now, it's winning deals.

Meta has historically been the poster child for open infrastructure: Open Compute Project, custom switching, disaggregated everything.

Going full-stack NVIDIA, from silicon to network fabric, signals that the performance and integration requirements of frontier AI training have overtaken the open-source hardware philosophy.

NVIDIA's push into Ethernet networking has previously faced scepticism from competitors and operators loyal to them. Meta validating Spectrum-X at hyperscale changes that conversation overnight.

We covered the Radian Arc acquisition last week (Issue #89) and flagged that the real question was how much of Submer's 5GW pipeline converts to operating capacity. Two signed deals within days of closing suggest the acquisition came with a warm pipeline and provide some support to the 5GW claim.

The India strategy is embed-and-expand, not build-and-own. Both partnerships deploy GPU infrastructure inside someone else's network: GTPL's telecom footprint and Datasamudra's colocation facility.

That keeps capex light while giving Submer sovereign compute positioning in a market where data residency requirements are tightening. India's government has been increasingly explicit about wanting AI workloads processed domestically.