AI Arms Race Just Went Into Overdrive
The AI Arms Race Just Went Into Overdrive — and Three IPOs Will Decide Who Pays for It
The week of April 14–21, 2026, delivered the most concentrated burst of competitive moves in the history of commercial AI. Three companies — Anthropic, OpenAI, and xAI — shipped major product releases within days of each other, each one calibrated not just for users, but for the investors who will soon decide the outcome of the largest technology IPOs ever attempted.
This is no longer a research race. It is a revenue race, a distribution race, and an IPO-timing race — running simultaneously.
April 16: Two flagship launches, one message
On April 16, Anthropic released Claude Opus 4.7, its most capable generally available model. The same day — not a coincidence — OpenAI shipped a sweeping update to Codex, transforming it from a coding assistant into a general-purpose AI workspace with desktop control, a built-in browser, image generation, and over 90 plugins.
Opus 4.7 leads the SWE-bench Pro coding benchmark at 64.3%, ahead of GPT-5.4’s 57.7%. It processes images at over three times the resolution of its predecessor. On paper, it is a strong release.
In practice, reception has been mixed.
The model uses a new tokenizer that can consume up to 35% more tokens per task depending on content — same price per token, higher effective cost per job. Temperature, top_p, and top_k parameters are locked to defaults; setting any non-default value returns an error. A Reddit thread titled “Opus 4.7 is not an upgrade but a serious regression” hit roughly 2,300 upvotes in 48 hours. Users accustomed to Opus 4.6’s conversational flexibility found the new model more literal, more deterministic, and less willing to infer intent from ambiguous prompts.
The complaints are real, but they miss the point. Opus 4.7 was not built for casual conversation. It was built for Claude Code — Anthropic’s agentic coding environment — where precise task execution, long-horizon reliability, and deterministic behaviour are features, not bugs. The model is tuned for enterprise workflows: well-specced tasks, structured inputs, measurable outputs. It is, in short, tuned for revenue.
Anthropic’s annualised revenue hit $30 billion by end of March 2026, up from $9 billion at end-2025 — a 3.3× increase in a single quarter, driven overwhelmingly by enterprise demand. More than 1,000 business customers now spend over $1 million annually. That growth is what makes the IPO possible, and Opus 4.7 is the model designed to sustain it.
OpenAI’s same-day Codex launch was equally strategic. The update added background computer use on macOS, letting agents operate desktop apps while the user works in parallel. Three million developers use Codex weekly; within ChatGPT Business and Enterprise, Codex usage grew 6× between January and April 2026. The message to enterprise buyers was clear: if you are already on ChatGPT, your coding tool is already here. No vendor switch required.
OpenAI’s missing flagship
The one thing OpenAI did not ship on April 16 was the model everyone is waiting for.
“Spud” — OpenAI’s internal codename for its next frontier model — completed pretraining on March 24, 2026, at the Stargate data centre in Abilene, Texas. Sam Altman confirmed it publicly. Greg Brockman called it the result of “two years of research” with a “big model feel.” The open question is whether it ships as GPT-5.5 or GPT-6 — a distinction that depends on how large the benchmark improvement is over GPT-5.4.
The release window was originally projected for mid-April. As of April 22, it has not shipped. Polymarket’s probability of a release by April 30 has dropped from roughly 78% to around 45%. The model appears to be in extended safety evaluation and red-teaming. OpenAI filled the gap with GPT-5.4-Cyber, a medical model called GPT-Rosalind, the major Codex refresh, and a new image generation model — gpt-image-1.5, which pairs a reasoning model with an improved generation engine to produce near-perfect images with consistent subject handling across poses and precise multi-prompt editing. These are real products. But they are side content. The flagship is still in the lab.
Meanwhile, OpenAI also shut down Sora, its video generation product, and ended a partnership with Disney — redirecting over 100,000 GPUs to Spud’s development. The resource allocation tells you where the company thinks the value is.
Musk’s $60 billion shortcut
On April 21, SpaceX announced a deal with Cursor — the most widely used AI coding environment — that includes an option to acquire the company for $60 billion later this year, with a $10 billion breakup fee if the deal falls through.
The move is a direct response to a competitive gap. xAI’s Grok models trail Anthropic’s Claude and OpenAI’s GPT series on coding benchmarks. Two of Cursor’s most senior engineering leaders, Andrew Milich and Jason Ginsberg, left to join xAI in March, reporting directly to Musk. SpaceX had already absorbed xAI in an internal merger in February 2026, valuing the combined entity at $1.25 trillion. An IPO roadshow is planned for the week of June 8, targeting a valuation of up to $1.75 trillion.
Neither Cursor nor xAI has proprietary models matching Claude or GPT. Cursor still sells access to both. The SpaceX deal is an attempt to solve xAI’s distribution problem by acquisition rather than model development — bundling Cursor’s 3+ million weekly developers into the SpaceX-xAI conglomerate ahead of the IPO. This is not a technology decision. It is a valuation decision.
The IPO collision course
What makes this moment structurally unusual is that three of the most valuable private companies in history are all preparing to go public within months of each other.
SpaceX (including xAI) is targeting a June 2026 roadshow at a valuation of up to $1.75 trillion, raising approximately $75 billion. OpenAI is planning a Q4 2026 listing at a target valuation of roughly $1 trillion, following a $122 billion private funding round that valued it at $852 billion. Anthropic, valued at $380 billion after its $30 billion Series G, has brought in Airbnb’s former IPO CFO and is targeting an October 2026 listing.
Combined, these three companies would need public markets to absorb somewhere between $2.9 trillion and $3.5 trillion in market capitalisation. For context: from 2016 to 2025, the entire US IPO market raised $469 billion in total. At a standard 15% float, these three IPOs alone would require $432–576 billion in new capital — roughly equivalent to the entire decade’s IPO proceeds, concentrated in a single quarter.
This is not normal. The sequencing and timing of these IPOs is as much a competitive variable as the models themselves. Each company wants to go public while its narrative is strongest and before the others absorb available capital. SpaceX going first in June puts pressure on OpenAI and Anthropic to either accelerate or differentiate. Every model release, every product launch, every benchmark in the next six months is partly an IPO marketing event.
What does a €500K-budget institution do with this information?
The immediate instinct might be: nothing. The IPO race is a Silicon Valley story. But for public institutions evaluating AI tools or planning digital infrastructure, the dynamics matter.
First, pricing is unstable. Every model provider is currently prioritising revenue growth over margin stability. Enterprise discounts, free tiers, and aggressive onboarding exist because these companies need ARR numbers that justify their valuations. That will change after the IPOs. Institutions locking in contracts now may get better terms than those negotiating in 2027.
Second, vendor lock-in risk is rising. The SpaceX-Cursor deal, the Codex super-app strategy, and Anthropic’s Claude Code ecosystem are all designed to make switching expensive. The A2A interoperability protocol (now at 150 organisations in production under the Linux Foundation) and Anthropic’s MCP are the closest things to interoperability standards — but they are young, and the incentive for dominant platforms to stay compatible weakens after IPO.
Third, the models are diverging. Opus 4.7 is tuned for structured enterprise work, not open conversation. Codex is becoming a desktop control surface, not a coding assistant. GPT’s image model now includes a reasoning engine. These are not interchangeable products. The “which AI should we use?” question is becoming “which workflow does each AI serve?” — and institutions that treat AI procurement as a single-vendor decision are already behind.
Finally, regulatory timing matters. The EU AI Act’s high-risk enforcement begins August 2, 2026 — between the SpaceX and OpenAI IPOs. Any institution deploying AI in healthcare, employment, or public services will need to demonstrate compliance with a framework whose harmonised standards are not yet published, whose enforcement contacts are not yet designated in most member states, and whose regulatory sandboxes are not yet operational. The companies racing to IPO are building products faster than the regulatory infrastructure can absorb them.
The race is not about who has the best model
The companies leading the AI race in April 2026 are not competing primarily on model quality. They are competing on distribution, on enterprise revenue, and on IPO timing. The models are necessary but not sufficient — what matters is who has the installed base, the enterprise contracts, and the narrative when the public markets open.
For institutions, the takeaway is straightforward: the AI market is entering its most volatile phase since the ChatGPT launch. Vendor positions will shift. Pricing will change. Product roadmaps will be driven by IPO narratives as much as by user needs. The institutions that will navigate this well are the ones that baseline their current state, understand what problem they are actually solving, and avoid locking into a single vendor before the market settles.
The ones that will not navigate it well are the ones buying a €500K platform because a vendor’s benchmark looked impressive this week.
DIGIPART helps public institutions assess where they stand before making technology decisions. We baseline current state, define phased roadmaps with named ownership, and ensure delivery holds after go-live — regardless of which vendor’s stock price is making headlines. If you are evaluating AI tools or planning digital infrastructure, talk to an advisor.
Sources: Bloomberg, Slashdot/The Verge, Evolink AI, FindSkill, FindSkill (Spud tracker), TechCrunch, CNBC, Tomasz Tunguz, Built In, PYMNTS, The Decoder
