AI Weekly #18/2026: Microsoft and OpenAI End Exclusivity – and Defuse the AGI Bomb

TL;DR

This week in 30 seconds:

• Microsoft × OpenAI: The exclusivity ends – OpenAI may now offer its models on AWS and Google Cloud, and the AGI clause (which would have terminated Microsoft’s access upon an AGI declaration) has been removed entirely. Microsoft retains access, loses control.
• AI Arms Race: The Pentagon has signed AI contracts with Nvidia, Microsoft, AWS, Google, OpenAI, and SpaceX – for classified networks up to Top Secret. Anthropic was left out after insisting on guardrails against autonomous weapons.
• Anthropic Valuation: The likely final private funding round before IPO: $50B in fresh capital, targeting ~$900B valuation – surpassing OpenAI’s $852B. A doubling in three months from $380B (Feb. 2026, per TechCrunch).
• Infrastructure Race: Alphabet, Amazon, Meta, and Microsoft will collectively spend $700B on AI infrastructure in 2026 – up 71% from 2024. Q1 alone: $130B combined.

---

Audio Version

14:23 | Download MP3

Chapters

- 0:00 - TL;DR - 1:16 - Story of the Week - 4:04 - More Top Stories - 8:17 - Quick Hits - 9:43 - Tool of the Week - 10:52 - Fail of the Week - 12:03 - Number of the Week - 12:45 - Reading List - 13:34 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

The End of OpenAI Exclusivity – and What the AGI Clause Really Meant

Since 2019, the partnership between Microsoft and OpenAI has been a quiet power arrangement: Microsoft pours in billions, receives exclusive cloud access in return, and – in one of the most remarkable contract clauses in corporate history – the right to automatically take over OpenAI’s technologies the moment OpenAI internally declared AGI [1]. That clause is now gone. Without replacement.

The renegotiated partnership, officially announced last week, fundamentally transforms the relationship: Microsoft loses exclusivity [1]. OpenAI may now distribute its products on any cloud provider – AWS, Google Cloud, Oracle Cloud Infrastructure, or directly via its own infrastructure. The $13B+ investment Microsoft put into the partnership [1] still buys access to the technology, but no longer control over its distribution.

The AGI clause is the real signal. For years it was the invisible governance problem lurking in the background: Who at OpenAI decides that AGI has been reached? And what happens to the board, the non-profit structure, the employees – when suddenly Microsoft holds first right to everything? The clause structurally constrained OpenAI’s strategic freedom, because every step toward genuine AGI capability simultaneously brought the “emergency exit” for Microsoft closer [1]. With its removal, the question arises anew: Who now defines AGI, and with what consequences for governance and regulation?

“OpenAI will now be able to distribute its products and models to additional cloud providers.” [1]

The renegotiation gives OpenAI the strategic freedom it needs for a multi-cloud strategy – and removes a fundamental conflict of interest around what AGI even means. For Microsoft, operationally little changes: Access to OpenAI models for Azure remains, Azure stays preferred cloud partner. But the balance of power has shifted.

For enterprises and developers, this means in the medium term: OpenAI products will be available on more infrastructures, price negotiations with a no-longer-exclusive provider will look different, and competition between cloud providers for OpenAI workloads can begin. The quiet winners: AWS and Google Cloud – both can now compete for OpenAI workloads that previously had only one infrastructure option.

Bottom Line: Microsoft invested $13B+ and still gets preferred access – but no longer control. OpenAI has shed its golden handcuffs. The real question is whether this resolves AGI governance problems or merely relocates them.

---

More Top Stories

Pentagon Arms Up: AI on Classified Networks – Anthropic Left Out

The US Department of Defense signed agreements last week with eight tech companies: Nvidia, Microsoft, AWS, Google, OpenAI, SpaceX, Reflection AI, and others [3]. The goal is deploying AI on classified networks at IL-6 (Secret) and IL-7 (Top Secret) levels – the highest available security classification for cloud infrastructure. The Pentagon states its objective explicitly: “United States military as an AI-first fighting force” [3].

Notable for who is absent: Anthropic. The company insisted on guardrails that would exclude use of the technology “for domestic mass surveillance and autonomous weapons” [3]. The Pentagon did not accept these conditions.

The contrast with OpenAI and Google is striking: both joined the Pentagon deals without comparable public conditions. Whether Anthropic’s position is primarily ethical or strategic in motivation cannot be definitively assessed from the outside – in effect, it means short-term exclusion from one of the largest military AI programs in the world. For enterprise decision-makers and compliance teams, this is relevant: Anthropic is positioning itself with publicly communicated use restrictions, which in regulated and security-critical environments can be both a differentiator and an obstacle.

---

Anthropic on Track for a $900 Billion Valuation – Before the IPO

Anthropic is expected to close a funding round of approximately $50B in fresh capital within two weeks [4]. The targeted post-money valuation is ~$900B – and could be exceeded due to high investor demand. For comparison: as recently as February 2026, Anthropic was valued at $380B [4]. In three months, the valuation has risen 137% [4].

This would put Anthropic ahead of OpenAI, which is currently valued at $852B post-money after its own $122B round earlier in 2026 [4]. The 48-hour allocation window for investors signals how hot the demand is. Some early 2024 investors are pausing and waiting for IPO liquidity – the round is considered the likely last before a 2026 IPO [4].

The number feels abstract until you put it in context: on paper, Anthropic is worth more than BMW, Deutsche Telekom, or BASF combined, without owning any hardware or physical infrastructure. The capital is meant to cover massive computing costs. The central question for 2026 and 2027 remains whether revenue growth from API, subscriptions, and enterprise licenses can justify the valuation over the medium term.

---

$700 Billion for AI Infrastructure – the Largest Investment Program in Corporate History

Alphabet, Amazon, Meta, and Microsoft will collectively spend nearly $700B on AI infrastructure in 2026 [2]. That is an increase of 71% over 2024 ($410B) – and the end of this growth curve is not in sight. In Q1 2026 alone, over $130B was spent combined [2].

The numbers behind the numbers: Meta’s Hyperion data center in Louisiana costs $27B and will house millions of GPUs [2]. A single Nvidia GPU costs up to $40,000 – and demand continues to outpace supply [2]. Investor reactions are mixed: Meta’s stock fell after Q1 earnings despite strong AI investment, while Alphabet and Amazon rose on cloud growth.

The paradox: the four companies are simultaneously building infrastructure for their own models and for competitors. AWS and Google Cloud sell compute capacity to Anthropic and OpenAI – the very companies that could potentially replace their own AI services. This is structurally unstable terrain, but for now infrastructure build-out buys time and market position.

---

Quick Hits

Briefly noted:

• Musk vs. OpenAI – a revealing admission: Elon Musk acknowledged in court testimony that xAI performed “distillation” on OpenAI models – meaning Grok was partly trained on outputs from GPT models [5]. More telling: Musk ranked Anthropic as the strongest AI provider (#1), ahead of OpenAI and Google – and described xAI as “a much smaller company with a few hundred employees” [5]. The man suing OpenAI for billions trained his own model on OpenAI outputs.
• Oscars Without AI Performances: The Academy has adopted binding rules for the 99th Oscars (2027): performances must be delivered by real actors (with their consent), screenplays must be demonstrably human-authored [6]. AI tools may still be used in a supportive capacity – but creative authorship must be human. The Academy can request disclosures about AI usage [6].
• Claude Connectors for Creative Software: Anthropic launches direct integrations for Blender, Autodesk Fusion, Adobe Creative Cloud (50+ tools), Ableton, Splice, SketchUp, Affinity by Canva, and Resolume Arena [7]. Claude can act as an on-demand tutor for complex software, write scripts, automate batch tasks, and translate formats. Educational partnerships with RISD, Ringling College, and Goldsmiths [7].

---

Tool of the Week

Stripe Link for AI Agents – Payment infrastructure for the agentic commerce era

Stripe has opened its Link wallet to autonomous AI agents: agents can process purchases, bookings, and tickets on behalf of users [8]. The permission model is well-designed – users authorize the agent via OAuth, the agent creates a purchase request, the user approves before payment. Agents never see raw payment data; instead, virtual cards with real-time authorization via “Issuing for agents” are provided [8].

For developers who want to equip agents with real purchasing capabilities, this is the most serious infrastructure solution to date. 90-day purchase protection with select merchants, configurable spending limits coming in the future, and stablecoin support in development [8]. Stripe is positioning itself as the payment rail for the agentic economy, before demand explodes.

Particularly relevant for teams building booking, shopping, or service agents: Stripe Link solves the “who pays and how?” problem with a framework that already accounts for users, liability, and compliance.

---

Fail of the Week

Sam Altman Criticizes Anthropic’s Cyber Restrictions – Then Introduces Exactly the Same Thing

A few weeks after Sam Altman called Anthropic’s decision not to publicly release its cybersecurity model Mythos “fear-based marketing,” OpenAI introduced exactly the same model for GPT-5.5: restricted access via “Trusted Access for Cyber (TAC)” [9]. Users must submit credentials and intended use before gaining access to penetration testing, vulnerability exploitation, and malware reverse engineering features. OpenAI is working with the US government to identify legitimate cyber defenders [9].

The wording of Altman’s original criticism of Anthropic – “fear-based marketing” – now hangs in the air while OpenAI applies the same decision logic he had criticized. Anthropic has surely noticed the screenshot potential of this moment.

What we learn: Once a company understands what powerful cybersecurity models can actually do, everyone reaches the same conclusion – no matter how loudly they previously said the opposite. The restriction is substantively justified. The communication strategy before it was not.

---

Number of the Week

$700 Billion – Big Tech’s AI Infrastructure Budget for 2026

Four companies – Alphabet, Amazon, Meta, Microsoft – are investing nearly $700B in AI infrastructure in a single year [2]. That is +71% compared to $410B in 2024. Q1 2026 alone: over $130B combined. For reference: the German federal budget for 2026 amounts to roughly $380B – four US tech giants are spending nearly twice as much on a single technology, in a single year. And no one knows where the curve ends.

---

Reading List

For the weekend:

1. The Next Phase of Our Partnership with Microsoft – OpenAI’s official statement on the renegotiation; worth reading for what is not said – about the AGI clause and what its removal means (6 min)
2. Pentagon Inks Deals with Nvidia, Microsoft and AWS to Deploy AI on Classified Networks – The full TechCrunch article on the Pentagon deal; essential for anyone who wants to understand Anthropic’s positioning and the implications for military AI governance (7 min)
3. Big Tech Hyperscalers Will Spend $700 Billion on AI Infrastructure This Year – Fortune analyzes why investors are reacting with mixed signals despite $700B CapEx, and what that means for cloud margins (9 min)

---

Next Week

What’s coming:

• Anthropic Round: The ~$900B funding round is reportedly set to close within two weeks [4]. If the valuation actually surpasses OpenAI’s, it will shift the entire industry’s valuation discourse.
• Pentagon AI in Practice: Following the signed framework agreements, first concrete deployment details at IL-6/IL-7 level are expected to emerge – and Anthropic’s response to its exclusion will become clearer.
• OpenAI TAC Program: How the “Trusted Access for Cyber” verification program works in practice, and whether it can genuinely distinguish legitimate cyber defenders from attackers, will become apparent in the coming weeks.

---

🤖 Behind This Newsletter

Generated in: ~40 minutes
Sources scanned: 9 articles from 5 feeds
Stories found: 11 → 9 selected
Validation: 4 agents, 8 issues found (3 blocking, 5 non-blocking) → 5 corrections applied
Model: Claude Sonnet 4.6 + Haiku (Validation)
Images: Pollinations.ai (1 generated)

Full Metrics

Phase	Metric	Value
Source Collection	RSS Feeds	5
Source Collection	WebSearch Queries	4
Selection	Stories presented	11
Selection	Stories selected	9
Draft	Words	~1,700
Draft	Sources cited	9
Validation	Fact-Check Issues	3
Validation	Balance Issues	3
Validation	Quality Issues	2
Validation	Legal Issues	0
Correction	Corrections applied	5

---

This newsletter was researched and written AI-assisted. Images generated with Pollinations.ai.