Five days. Five stories that will still matter in five years. Anthropic filed to go public at a valuation that makes it the most valuable AI startup in history. OpenAI overhauled how ChatGPT remembers you. SoftBank committed €75 billion to build AI data centres in France. Microsoft showed us what AI-powered Windows looks like. And NVIDIA moved its most powerful hardware off the server rack and onto the desk.
This week in AI (June 2–6, 2026): Anthropic confidentially filed for an IPO at a ~$965 billion valuation with a $47B revenue run rate. OpenAI launched Dreaming V3, a rebuilt ChatGPT memory system that synthesises context automatically from past conversations. SoftBank announced a €75 billion commitment to build 5GW of AI data centre capacity in France. Microsoft Build 2026 unveiled Windows Agent Framework APIs and a $9.69B US DoD contract. NVIDIA announced RTX Spark, a desktop PC with up to 128GB unified memory for running AI models locally.
This is not a slow news week. It is the week AI stopped being a technology story and started being an everything story — infrastructure, finance, memory, operating systems, and geopolitical capital allocation all in the same five-day window.
1 Anthropic Files for an IPO — and the Numbers Are Staggering
What happened
On June 1, 2026, Anthropic confidentially submitted a draft S-1 registration statement to the US Securities and Exchange Commission, officially starting the process for an initial public offering. The headline figure: a $65 billion Series H round placed the company’s post-money valuation at approximately $965 billion — the highest ever for a private AI company, ahead of OpenAI’s $852 billion valuation.
The revenue numbers are equally striking. Anthropic’s annualised revenue run rate hit roughly $47 billion in May 2026, up from approximately $10 billion the prior year — nearly a 5x growth rate in twelve months. Goldman Sachs, JPMorgan Chase, and Morgan Stanley are handling the deal.
What it means
Anthropic joins OpenAI (which filed confidentially around May 22) and SpaceX in what could become one of the most consequential IPO windows in decades. If all three list as expected in fall 2026, more than $200 billion in new public market value could arrive at once — in a year when the entire US IPO market raised only $45 billion.
Going public changes Anthropic’s accountability structure permanently. Every quarter, it will have to explain its compute costs, safety research spending, and margins to public shareholders. The tension between commercial scale and the company’s stated mission of beneficial AI becomes a very public conversation. That transparency has no precedent in frontier AI lab history.
There is also an infrastructure signal worth noting: Anthropic currently pays SpaceX $1.25 billion per month to use compute at SpaceX’s Colossus 1 data centre in Memphis. That figure, visible in SpaceX’s own prospectus, illustrates the capital intensity of running frontier AI at scale.
Honest caveat: A confidential filing is not a commitment to go public. Anthropic has said the filing “gives us the option” to list after SEC review. No price, date, or share count has been set. The direction of travel is clear — but this is not a done deal.
2 ChatGPT Gets a Memory Upgrade — and It’s More Consequential Than It Sounds
What happened
On June 4, 2026, OpenAI began rolling out Dreaming V3 — a rebuilt memory architecture for ChatGPT — to Plus and Pro users in the United States. Free and Go users are expected to receive it within weeks. The old system worked like a notepad: you had to explicitly tell ChatGPT to remember something. Dreaming V3 runs a background process after conversations end, automatically synthesising preferences, constraints, ongoing projects, and recurring context — without you asking.
Factual recall improved from 67.9% to 82.8%. Preference adherence improved from 55.3% to 71.3%. Accuracy over time improved from 52.2% to 75.1%. The update also introduces a readable memory summary page where users can see what ChatGPT has stored, edit it, and configure which topics it should raise.
What it means
People stop using AI products not because they can’t solve hard problems, but because the AI keeps asking them to repeat themselves. Dreaming V3 is fundamentally a relationship repair feature — it makes ChatGPT feel like it pays attention across sessions, not just within them.
Memory is becoming the differentiator. The chatbots that feel most useful over time will be the ones that don’t start from zero every conversation.
— NeeAr Ventures Editorial
A February 2026 academic paper analysed 2,050 memory entries from 80 ChatGPT users and found that 96% were created by the system unilaterally — without user prompting. An AI that infers your behavioural profile from your conversations will face growing scrutiny. The EU AI Act’s transparency obligations take effect in August 2026, and OpenAI’s international rollout falls right before that deadline.
3 SoftBank Bets €75 Billion on Europe — and Picks France as the AI Continent
What happened
On June 1, 2026, at France’s annual Choose France summit at the Palace of Versailles, SoftBank Group announced it would invest up to €75 billion to build and operate 5 gigawatts of AI data centre capacity in France. The first phase: €45 billion for 3.1 GW of capacity across three new data centres in northern France (Dunkirk, Bosquel, and Bouchain), targeted for completion by 2031.
This was the single largest investment at a summit that attracted €93 billion in total foreign investment commitments — the highest total in the Choose France summit’s history. Macron said the projects would create 15,600 jobs in a country where unemployment sits at 8%.
What it means
Three things make this globally significant, not just European.
For two years, the narrative was that AI infrastructure was concentrated in the United States and China, with Europe struggling to keep pace. SoftBank’s commitment — combined with a separate €50 billion UAE–France AI data centre deal announced the same week — signals that capital is now actively flowing into European AI infrastructure at a scale not visible twelve months ago.
Europe’s high energy costs have been a documented barrier to data centre competitiveness. France’s nuclear-powered electricity grid makes it structurally cheaper to run compute-intensive workloads than most European alternatives. This is showing up in investment decisions, not just policy documents. SoftBank’s shares rose 14% on the announcement day and are up over 70% in 2026 so far.
SoftBank’s investments are concentrated in Arm Holdings (whose chip designs power AI servers globally), OpenAI, and now direct data centre infrastructure. When it commits €75 billion to a single country, it is pricing in a multi-decade AI infrastructure supercycle — not a short-term bet.
Question to track: Europe’s regulatory environment (GDPR, EU AI Act, forthcoming transparency obligations) is significantly more complex than the US. Whether that overhead affects the economics of these data centres — or becomes a compliance-by-design selling point for enterprises — will play out over the next three to five years.
4 Microsoft Build 2026 — AI Agents Come to Windows
What happened
Microsoft held Build 2026 in San Francisco the week of June 2. The key announcements included: homegrown MAI models for GitHub Copilot (directly addressing the backlash over Copilot’s new token-based billing that took effect June 1), Windows Agent Framework APIs for OS-level autonomous agents, Copilot Agent Mode for multi-step GitHub coding workflows, Azure AI Foundry formally adding Anthropic Claude, and a $9.69 billion DoD enterprise contract for Microsoft 365, Azure, and AI Copilot services across approximately 2.8 million US Department of Defense staff — the largest single Microsoft government contract in company history.
What it means
The operating system is becoming an AI platform. Windows Agent Framework APIs mean developers can now build agents that operate at the OS level, executing tasks across applications — not just inside a browser tab or side panel.
On June 1, GitHub Copilot switched from flat-subscription to token-based billing. One developer reported that a $29/month subscription could become $750/month under the new model. The backlash was immediate and loud on Reddit, Hacker News, and X. Microsoft’s announcement of homegrown coding models at Build was, in part, a direct response — a signal that they understand the pricing problem and are working on it from the model cost side.
5 NVIDIA RTX Spark — Frontier AI Off the Server Rack and Onto Your Desk
What happened
NVIDIA announced RTX Spark at Microsoft Build — a desktop PC with up to 128GB of unified memory and CUDA support, designed specifically for running AI models locally on Windows. Paired with Nemotron 3 Ultra (NVIDIA’s open enterprise model released the same week), the announcement signals a push to bring serious AI inference capacity out of the cloud and into local environments. Systems are expected to ship in fall 2026.
What it means
The AI infrastructure narrative for two years has been: models run in the cloud, you access them through APIs. RTX Spark points at a different trajectory — where capable models run on your hardware, in your building, without per-query cost or cloud dependency. For privacy-sensitive industries (legal, healthcare, government, financial services), local inference has been the missing piece. This could be it.
The question fall 2026 will answer: What are the actual prices, thermals, and buyer demand when systems ship? The announcement is compelling; the market test is still ahead.
Quick Hits This Week
- Project Glasswing expands: Anthropic extended its program giving vetted organisations access to Claude Mythos Preview for cybersecurity to cover critical infrastructure sectors (June 2). Claude now connects to 28 security platforms through the Claude Compliance API.
- Claude Sonnet 4.8 rumours: Speculation about an upcoming Claude Sonnet 4.8 model continued gaining credibility this week. Anthropic has made no announcement — treat as rumour.
- AI model benchmarks: Google’s Gemini 3.5 Flash and OpenAI’s GPT-5.5 Instant are both positioned as fast, cost-efficient options; Anthropic’s Claude Opus 4.8 is generating enterprise momentum from independent benchmarks.
- AI training costs compressing: Orion-100B trained a 100-billion-parameter model for $1.25/hour — a data point in the ongoing trend of frontier model training becoming dramatically cheaper over time.
What This Week Actually Tells Us
Take a step back and look at the shape of the week. A company builds AI. It raises money. It goes public. It pays for compute by the billion per month. It competes. Governments decide what rules apply. The hardware gets cheaper and more powerful. The models get embedded in operating systems and enterprise tools. And capital flows across continents to wherever the infrastructure can be built.
That is the full loop, and this week had visible pieces of every part of it. The SoftBank story and the NVIDIA story — the ones that get less attention than the IPO — will matter most over the next decade. Legislation moves slowly until it doesn’t. Hardware that runs frontier models locally changes the economic model of AI deployment entirely.
The Anthropic IPO will dominate headlines. The others deserve your attention too. Stay curious.
Anthropic confidentially filed for an IPO valuing it at approximately $965 billion. OpenAI upgraded ChatGPT’s memory with Dreaming V3, a system that synthesises context from past conversations automatically. SoftBank committed €75 billion to build AI data centres in France. Microsoft Build 2026 showcased Windows Agent Framework APIs and a $9.69 billion DoD contract. NVIDIA announced RTX Spark, a desktop PC with up to 128GB of unified memory for running AI models locally.
Anthropic’s confidential S-1 filing on June 1, 2026 values the company at approximately $965 billion, making it the most valuable AI startup in history, ahead of OpenAI’s $852 billion private-market valuation. The company’s annualised revenue run rate hit $47 billion in May 2026, up roughly 5x from $10 billion the prior year. Combined with OpenAI’s filing and SpaceX’s planned listing, the fall 2026 IPO window could bring over $200 billion in AI public-market value to investors at once.
Dreaming V3 is OpenAI’s rebuilt memory architecture for ChatGPT, launched June 4, 2026. Unlike the previous system that required users to explicitly save information, Dreaming V3 runs in the background after conversations end to automatically synthesise relevant context. According to OpenAI’s internal benchmarks, it improves factual recall from 67.9% to 82.8% and preference adherence from 55.3% to 71.3%. It rolled out first to Plus and Pro users in the US, with Free users to follow.
SoftBank announced on June 1, 2026 at the Choose France summit that it would build 5 gigawatts of AI data centre capacity in France, starting with €45 billion for sites in northern France. France’s nuclear-powered electricity grid gives it a structural energy cost advantage over most of Europe, making it attractive for compute-intensive AI infrastructure. The announcement was the single largest investment at a summit that attracted €93 billion in total foreign investment commitments.
Microsoft Build 2026 announced homegrown MAI models for GitHub Copilot, Windows Agent Framework APIs for OS-level autonomous agents, Copilot Agent Mode for multi-step coding workflows, Azure AI Foundry adding Anthropic Claude, and a $9.69 billion DoD enterprise contract covering approximately 2.8 million US Department of Defense staff — the largest single Microsoft government contract ever.