Editor's note. The cycle is again capital-markets-heavy, so the lead and one main sit in the Enterprise/markets bucket — flagged for honesty against the topic-balance guidance. The justification is that both are genuinely new and directly bear on what a European buyer pays and depends on: an emerging inference price war (lead) and the largest IPO on record actually trading today (main). The Open Source main is the open-weight cohort's overdue reproducibility test, not a release. Regulation is held to a single quick hit. No frontier model shipped in the last 24 hours; the freshest model news is a non-release (Gemini 3.5 Pro still owed) and a non-delivery (MiniMax M3 weights), both treated as such.
The first real AI price war is forming — and the buyer is the reason
OpenAI is weighing significant cuts to its token prices to keep customers from defecting to Anthropic, according to a Wall Street Journal report carried on 11 June by Bloomberg and CNBC. The deliberations are early and nothing is decided, but the framing is telling: OpenAI would cut if Anthropic cuts, in the run-up to two listings that will put both companies' economics in front of public-market investors. OpenAI confidentially filed its S-1 on Monday, days after Anthropic's own filing and two weeks after Anthropic closed its Series H at a $965bn valuation, just ahead of OpenAI's $852bn March mark. A token war is the predictable consequence of two near-trillion-dollar rivals competing for the same enterprise workloads while preparing to be judged on growth.
For a European deployer the price war reads two ways. The obvious read is good news: per-token costs, already down sharply, fall further, and the case for routing the easy majority of traffic to the cheapest adequate model only strengthens. The less obvious read is the durable one. Frontier pricing is becoming a competitive weapon set by two US companies that lose billions today — OpenAI's filings project an operating loss of roughly $74bn in 2028 alone — cumulative burn approaching $115bn through 2029 — with first profit targeted for 2029 — which means the number can move on strategy rather than cost, down to win share now and up once the market consolidates. A price you do not control is not a foundation you can plan on; portability and provider-agnostic routing are what let a buyer take the discount without inheriting the dependency.
The AI build-out lands on Nasdaq
SpaceX began trading on Nasdaq today under SPCX, priced at $135 a share for a roughly $75bn raise and a $1.75–1.8tn valuation — the largest listing on record. Demand was extraordinary: total orders topped $250bn and retail interest alone exceeded $100bn, with the book oversubscribed close to four times and up to 30% of shares reserved for retail, well above the usual single-digit allocation. The figure that matters for this brief is the AI demand underneath the prospectus: Google reportedly pays SpaceX around $920m a month for compute, and xAI ran a $6.4bn operating loss and $12.7bn of AI capex per the SpaceX prospectus against roughly $3.2bn of 2025 revenue.
The signal is not the rocket company; it is that the capital intensity of the AI frontier is now being underwritten by public equity. A small set of US-listed entities — compute landlords, model vendors, and now public companies answerable to quarterly demand — increasingly sits upstream of every enterprise's AI spend. Paired with the price-war lead, the same picture comes into focus from two angles: the cost of frontier AI and the structure that supplies it are both being shaped by US public markets. For a regulated European buyer the mitigations are unchanged and unglamorous — portability, the option to host open weights yourself, and a routing layer that is not bolted to any one provider's equity story.
The open-weight counterweight is late
MiniMax committed to publishing M3's open weights and a full technical report within about ten days of its 1 June launch, putting the deadline at roughly 11 June. As of today the weights are still absent from MiniMax's Hugging Face organisation and its M3 GitHub repository, more than eleven days on. The claims are what make the delay matter: 59% on SWE-Bench Pro — which MiniMax says edges GPT-5.5 and Gemini 3.1 Pro — a one-million-token context on a new sparse-attention design, and native multimodality at roughly a tenth of frontier API cost. None of it is independently verifiable until the weights and report ship and evaluators such as Artificial Analysis and LMArena can score them.
This is where the open frontier's credibility is actually tested. Against a forming price war among closed US labs, the open-weight pitch is precisely that a model you can download, audit and run in a jurisdiction of your choosing is the one that survives a contract renegotiation, a retention-clause change, or a strategic price hike. That argument only holds if the weights ship as promised, on time and reproducibly — and if the licence permits commercial use, which for MiniMax's earlier M2.7 required written authorisation and is worth reading closely. An overdue release is not a scandal, but it is a reminder that the counterweight only counts when it is on the scale.
Quick Hits
- EU AI Act clock, one regulation item — the Commission published a Code of Practice on marking and labelling AI-generated content on 10 June, fleshing out the Article 50 transparency duties. GPAI obligations are 51 days out (2 August 2026), and the Digital Omnibus high-risk postponement (Annex III to Dec 2027, AI-in-products to Aug 2028) still awaits Official Journal publication — provisional until it lands.
- Gemini 3.5 Pro still hasn't shipped — Google promised June GA at I/O (2M context, Deep Think; a ~$15/$60 per-million read), but the model remains in limited Vertex preview. With three weeks left in the month, the slip keeps Fable 5 ($10/$50) a clear window at the top of the price ladder.
- The open inference stack keeps pace — vLLM shipped 0.22.1 on 5 June and rolled out day-0 support for NVIDIA's Nemotron 3 Ultra a day earlier, so the engine most self-hosters run is current with the newest open-weight models the moment they appear — the practical underside of "host it yourself."
- The accounting becomes a competitive surface — as both labs head for public markets, OpenAI has accused Anthropic of inflating revenue by recognising gross rather than net amounts on partner contracts; Anthropic says it follows standard practice. The dispute is a preview of the legibility an IPO forces on numbers that have been private until now.