Capital & Industrial Strategy

SpaceX has closed a confirmed $6 billion compute agreement with Reflection AI, an Nvidia-backed open-source lab, for access to GB300 Nvidia chips across its Colossus 2 data centre near Memphis at $150 million per month from July 2026 through 2029. CNBC reports that SpaceX has also signed compute deals with Anthropic, Google, and Cursor, confirming Colossus 2 is now operating as a commercial neocloud. This is a structural shift: SpaceX is no longer merely an AI infrastructure consumer but a supply-side competitor to CoreWeave and Oracle Cloud Infrastructure, with the advantage of massive, already-deployed Nvidia capacity.

To fund expansion, SpaceX is launching its debut investment-grade US dollar bond offering in what Bloomberg describes as the start of a major borrowing spree following its $75 billion IPO. The combination of contracted compute revenue — providing predictable cash flows to service debt — and access to public bond markets gives SpaceX a materially different capital structure than private neocloud rivals. Masayoshi Son has publicly dismissed Elon Musk's orbital data centre concept, predicting the AI race will be won on terrestrial compute, which implicitly validates SpaceX's ground-based Colossus strategy while undercutting any premium investors might ascribe to speculative space-based infrastructure plays.

Groq has confirmed a $650 million raise following Nvidia's not-acqui-hire deal that pulled key personnel into the Nvidia orbit. TechCrunch reports the company is re-staffing executive positions and doubling down on its neocloud business — selling inference-as-a-service on its LPU hardware rather than competing for chip sales against Nvidia's distribution network. The strategic logic is clear: Groq's LPU architecture has a defensible inference latency advantage that is commercially valuable even if Groq cannot win the training chip market.

The raise confirms that investors still see a viable business in specialised inference infrastructure, even after a destabilising talent event. The neocloud framing — monetising proprietary silicon through cloud services rather than chip sales — mirrors what Cerebras and others are attempting. The critical question is whether Groq's inference performance advantage is durable as Nvidia's Blackwell and GB300 architectures close the latency gap on inference workloads.

Qualcomm is in advanced talks — confirmed by Bloomberg as a near-deal — to acquire Modular Inc. at approximately $4 billion. Modular builds AI infrastructure software, notably the MAX inference engine and the Mojo programming language, designed to make AI model deployment portable across heterogeneous hardware. The strategic intent is explicit: Qualcomm needs a software layer that makes its Snapdragon and AI chip portfolio attractive to enterprise developers who currently default to Nvidia's CUDA ecosystem. Bloomberg notes the deal is not yet closed and remains subject to finalisation.

This acquisition, if completed, would give Qualcomm a credible answer to CUDA lock-in — the primary barrier preventing enterprise AI workloads from migrating to non-Nvidia hardware. At $4 billion, it is a meaningful but not outsized price for a company whose software could theoretically unlock Qualcomm's entire on-device and edge AI hardware roadmap for enterprise customers. The deal also reflects a broader pattern of semiconductor companies vertically integrating into software to defend hardware market share.

The Economist reports that entrepreneurs, exchange operators, and AI firms are actively constructing tradable financial instruments backed by compute capacity — effectively creating a futures and spot market for GPU hours. This development sits at the intersection of commodity finance and AI infrastructure: if compute can be standardised and traded like power or bandwidth, it changes how capital is allocated to data centre buildout, how AI startups hedge their input costs, and how investors gain exposure to AI infrastructure without direct equity ownership.

The Reflection AI deal with SpaceX — $150 million per month in contracted compute — illustrates the commercial precursor: long-dated compute contracts that resemble take-or-pay commodity agreements. The next step, as The Economist frames it, is securitising or exchanging these contracts. This would have significant implications for AI startup financing, since compute access is currently the primary capital constraint for frontier model development, and a liquid compute market would allow startups to lock in capacity costs and present more predictable cost structures to equity investors.

Getty Images surged 145% after announcing a confirmed licensing deal with OpenAI, Bloomberg reports, resolving a major outstanding IP dispute and giving OpenAI access to Getty's commercial image library for training and generation. Simultaneously, IBM and OpenAI announced a confirmed enterprise security AI partnership, and OpenAI launched new creative tools targeting advertising — described by Semafor as a $100 billion market play. Together these moves show OpenAI executing a multi-front commercialisation strategy: licensing content inputs, partnering with enterprise incumbents for distribution, and building vertical-specific products.

Against this expansion, the Wall Street Journal reports that Sam Altman's personal investments in companies including Helion have seen material upswings correlated with OpenAI's commercial relationships with those same entities. WSJ presents this as a documented pattern rather than speculation. For institutional investors evaluating OpenAI's upcoming capital structure — including its ongoing conversion from nonprofit — governance risk and related-party transaction scrutiny are now material considerations, not peripheral concerns.