AI Weekly: 05/18/26
Cerebras goes public in the biggest tech IPO of 2026, OpenAI saves $97B in renegotiated Microsoft deal, and Anthropic surpasses OpenAI in business customers for the first time
Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Cerebras went public in the biggest tech IPO of 2026, raising $5.5 billion and seeing its stock pop 108% on its first day of trading to a $66 billion valuation. OpenAI’s renegotiated deal with Microsoft caps revenue-sharing payments at $38 billion, saving the company an estimated $97 billion through 2030, while Microsoft has already generated over $30 billion in revenue from OpenAI’s technology and spending.
Anthropic surpassed OpenAI in business customer adoption for the first time according to Ramp data, launched a new small business initiative with integrations into QuickBooks, Canva, and HubSpot, and saw its legal AI ambitions put Clio’s $500 million milestone in a new light. Meanwhile, BlackRock is weighing a $5 to $10 billion investment in SpaceX’s upcoming IPO at a potential $1.75 to $2 trillion valuation.
Also this week, Greg Brockman took charge of OpenAI’s product strategy while Fidji Simo remains on medical leave, Google and SpaceX are in talks to put data centers into orbit, and GM laid off hundreds of IT workers to hire those with AI skills instead.
More on this past week’s top AI headlines below!
- ZG
Here are the most important stories of the week:
AGENTS
Notion launched a developer platform that transforms its workspace into an orchestration hub for AI agents, letting users chat directly with external agents including Claude Code, Cursor, and Codex, assign them work, and track progress alongside human collaborators. Link.
New features include Workers (a cloud-based environment for running custom code in a secure sandbox), Database Syncing (pulling data from Salesforce, Zendesk, Postgres, and others directly into Notion databases), and MCP Protocol support for agent tools.
CEO Ivan Zhao said “any data, any tool, any agent, that’s the big picture for the Notion Developer Platform,” acknowledging that “historically, Notion hasn’t been the most developer-focused platform.”
Notion is positioning itself as an orchestration layer that coordinates AI work across multiple tools and data sources, with Decagon and additional partner agents planned beyond the initial launch.
Apple is designing a framework to allow AI agent apps into the App Store while maintaining its privacy, security, and payment rules, with potential updates coming at WWDC 2026 on June 8. Link.
The core challenge is that agents can dynamically spin up sub-applications after Apple has already approved the parent app, circumventing the review process and potentially bypassing App Store fees.
Apple’s staffers are building a system to prevent the kind of runaway behavior seen in agentic systems, including agents going haywire and deleting users’ emails or spreading malware.
The existing app review model, where Apple approves a static app, breaks down when an agent can generate new functionality on the fly after approval.
Anthropic launched “Claude for Small Business,” a new suite of services targeting local businesses like hardware stores and coffee shops, with integrations between Claude Cowork and QuickBooks, Canva, Docusign, HubSpot, and PayPal. Link.
The features are available via a newly introduced toggle within Claude Cowork, Anthropic’s task-automation platform that can browse the web, manage files, and execute multistep workflows on a user’s behalf.
Anthropic is planning a coast-to-coast promotional tour starting in Chicago and hitting 10 cities, offering free AI training workshops available to 100 local small business leaders at each stop.
The initiative represents a significant strategic shift for Anthropic, moving beyond enterprise and developer customers to target the massive small business market segment.
CODING/DEVTOOLS
OpenAI announced that Codex is being integrated into the ChatGPT mobile app, allowing developers to assign coding tasks, ask questions about their codebase, and monitor agent progress from their phones. Link.
The latest version runs on GPT-5.3 Codex, which is 25% faster than GPT-5.2 and was notably the company’s first model that “was instrumental in creating itself.”
Users can find the tool in ChatGPT’s sidebar, assign the agent new tasks by typing a prompt and clicking the “Code” button, and ask questions about their codebase via the “Ask” button.
The move intensifies competition with Anthropic’s Claude Code and is part of OpenAI’s broader push toward a unified AI “super app” combining ChatGPT, Codex, and its AI browser.
Canadian legal tech company Clio hit $500 million in ARR, but the milestone is complicated by Anthropic’s simultaneous expansion of “Claude for Legal” with new plug-ins and MCP connectors, turning a key supplier into a competitor. Link.
Clio was valued at $5 billion in its $500 million Series G last November and completed a $1 billion cash-and-stock acquisition of vLex, a 26-year-old legal data intelligence platform.
Both Harvey (which raised $200 million at an $11 billion valuation in March) and Legora (which raised a $600 million Series D) rely on Claude as a core model, creating an uncomfortable dynamic where their key AI supplier is now also a direct competitor.
Legora launched a high-profile ad campaign featuring actor Jude Law as the legal AI market rapidly heats up with multiple well-funded players.
SPEECH/AUDIO
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, announced a research preview of a “full duplex” AI interaction model that responds in 0.40 seconds and can listen and speak simultaneously, making conversations feel more like a phone call. Link.
TML-Interaction-Small roughly matches the speed of natural human conversation and is significantly faster than comparable models from OpenAI and Google; the “full duplex” capability allows for natural interruptions and back-and-forth.
This is a research preview, not a consumer product; Thinking Machines was founded in early 2025 and closed a $2 billion round at a $10 billion valuation (later valued at $12 billion).
The lab lost two co-founders back to OpenAI in January 2026 but deepened its ties with Google in April through a multibillion-dollar infrastructure deal.
RESEARCH
Junyang Lin, the architect behind Alibaba’s Qwen open-source model family, launched a new AI lab focused on world models and embodied intelligence, seeking hundreds of millions at a $2 billion valuation with backing from HongShan (formerly Sequoia China) and Gaorong Ventures. Link.
Lin publicly announced his resignation on X in March, posting “bye my beloved qwen”; during his three years at Alibaba he elevated Qwen into one of the world’s leading open-source AI model families.
A $2 billion valuation for a startup with no product yet is nearly unprecedented in China, where AI startups are typically valued far lower than U.S. peers; Lin’s personal track record with Qwen justifies the premium.
Lin has already recruited core team members from ByteDance, Tencent, and overseas AI companies, shifting his focus from large language models to robotic AI brains.
INFRASTRUCTURE
Cerebras went public in the biggest tech IPO of 2026, raising $5.5 billion at $185 per share before the stock surged 108% to close at $311 on its first day of trading, giving the AI chip maker a roughly $66 billion valuation. Link.
Revenue was approximately $510 million in 2025 (up 76% YoY) and the company swung to $237.8 million in net income after losing nearly half a billion the year before; Benchmark’s stake was worth $5.5 billion at close and Eclipse turned a $6.5 million Series A into a $2.5 billion return.
OpenAI signed a $10 billion compute deal with Cerebras and received warrants that could give it an ~11% equity stake, worth approximately $11.7 billion at the IPO opening price if fully vested; Sam Altman and Greg Brockman are also personal angel investors.
The IPO priced far above the initial $115 to $125 range (later raised to $150 to $160); co-founder and CEO Andrew Feldman’s stake was worth nearly $1.9 billion at the IPO price.
OpenAI’s renegotiated deal with Microsoft caps total revenue-sharing payments at $38 billion through 2030, saving OpenAI an estimated $97 billion compared to the previous uncapped arrangement, while shifting to a non-exclusive licensing model. Link.
Under the new terms, OpenAI pays Microsoft 20% of revenue with a hard cap at $38 billion; the old arrangement had no ceiling and was tied to the undefined question of when OpenAI achieved AGI.
Microsoft’s exclusivity to sell OpenAI’s models expired in April 2026, allowing OpenAI to offer on AWS and Google Cloud; Microsoft’s investment is now valued at approximately $135 billion, representing a 27% diluted ownership stake.
OpenAI paid Microsoft roughly $23 billion in Azure server rentals between 2023 and 2025; the new IP arrangement shifts to non-exclusive licensing running through 2032.
Microsoft has generated more than $30 billion in revenue from OpenAI’s technology and spending between 2023 and 2025, more than doubling its $13 billion investment, but has spent over $100 billion in infrastructure costs to support the partnership. Link.
The largest revenue component was OpenAI spending approximately $23 billion renting Azure servers, with the remainder from AI product sales including Office 365 Copilot, GitHub Copilot, and reselling OpenAI’s models to enterprise customers.
Microsoft was the only cloud provider allowed to sell OpenAI’s LLMs to businesses for three years, an exclusive right that expired in April 2026; capital expenditure is expected to exceed $40 billion in Q4.
While the $30 billion in revenue more than doubles the $13 billion investment, the $100 billion-plus in infrastructure spending means Microsoft is still significantly underwater on a pure cost basis.
OpenAI has structured massive purchasing commitments with suppliers like Cerebras and Nvidia that generate billions in financial value for OpenAI through equity stakes, warrants, and circular deal structures before any chips are even delivered. Link.
OpenAI committed $20+ billion to buy Cerebras chips over three years, receiving warrants that could give it an ~11% equity stake worth approximately $11.7 billion at the IPO opening price if fully vested; OpenAI also provided Cerebras a $1 billion loan for data center construction.
In its $122 billion funding round, Amazon committed up to $50 billion and Nvidia $30 billion, both major suppliers to OpenAI; Amazon extended an AWS deal to $100 billion over 8 years while OpenAI committed to buying 3 GW of Nvidia chips.
Sam Altman appeared in a video to promote Cerebras’ IPO, underscoring the intertwined relationship; Altman personally held ~89,000 Cerebras shares worth ~$16.5 million at the IPO price.
Google and SpaceX are in active discussions to build data centers in orbit, as SpaceX gears up for its anticipated $1.75 trillion IPO and pitches investors on the idea that space will become the cheapest location for AI compute within the next few years. Link.
Google has its own related initiative called Project Suncatcher with plans to launch prototype satellites by 2027; Mike Safyan, an executive at Planet Labs (building the prototypes for Google), confirmed the plans.
The broader industry is moving fast: Blue Origin entered the space data center market in March 2026, Cowboy Space raised $275 million to build rockets specifically for this purpose, and SpaceX has sought federal approval to launch up to 1 million solar-powered satellite data centers.
The economics remain brutal, as terrestrial data centers are still far cheaper once satellite construction and launch costs are factored in.
BlackRock is in discussions to invest $5 to $10 billion in SpaceX’s upcoming IPO from its $536 billion in actively managed funds, as SpaceX targets a valuation of $1.75 to $2 trillion and aims to raise approximately $75 billion in what would be the largest stock-market debut in history. Link.
SpaceX targets June 11 pricing and June 12 listing on Nasdaq under ticker “SPCX,” with a roadshow launching June 4; the timeline accelerated due to faster-than-expected SEC review.
The IPO would dwarf Saudi Aramco’s $29 billion 2019 offering; SpaceX merged with xAI in February 2026, so investors are effectively buying into a combined space, rocket, and AI business.
Both BlackRock and SpaceX declined to comment; the final investment amount could change based on IPO pricing.
Modal, a cloud infrastructure startup for AI developers, is in talks to raise $150 to $250 million at a $4.5 billion valuation after its revenue surged fivefold to approximately $300 million in annualized run rate. Link.
The $4.5 billion target is an 80% increase from its last round; Modal was in talks to raise at $2.5 billion as recently as February 2026.
Modal rents out Nvidia GPUs and provides software to help developers run, train, and deploy AI models and agents; key customers include Ramp and Lovable.
The key growth driver is rising enterprise demand for AI agent sandboxes, isolated environments where companies can safely develop and test autonomous agents.
ENERGY
Elon Musk’s xAI is operating nearly 50 natural gas turbines at its Mississippi data center (Colossus 2) without proper regulatory oversight, exploiting a loophole that classifies turbines on flatbed trailers as “mobile” to dodge air pollution regulations. Link.
Of 59 turbines on-site, only 15 have been properly permitted and 18 are classified as temporary, meaning regulators do not track their pollution; the NAACP has filed a lawsuit on behalf of local residents arguing the unchecked emissions are worsening air quality.
The Southern Environmental Law Center estimates xAI has operated at least 35 turbines capable of emitting more than 2,000 tons of NOx pollution, which contributes to smog and respiratory problems.
This follows the EPA’s January 2026 ruling that xAI’s natural gas generators at its Memphis facility were illegally used; the NAACP asked the court this week for an injunction against xAI.
OTHER
Anthropic has surpassed OpenAI in business customer adoption for the first time, with 34.4% of participating businesses paying for Anthropic services versus 32.3% for OpenAI, according to expense data from fintech company Ramp covering more than 50,000 companies. Link.
The shift has been dramatic: in May 2025, only 9% of businesses were paying for Anthropic products, but that figure climbed roughly 26 percentage points over the following 12 months; over the same period, OpenAI’s share actually declined by 1%.
Anthropic’s annualized revenue jumped from $9 billion at the end of 2025 to $30 billion by the end of March 2026, driven largely by demand for Claude Code and related coding products.
The data comes as Anthropic reportedly could raise a new $50 billion round at a $900 billion valuation.
OpenAI co-founder and president Greg Brockman is taking over product strategy on an interim basis while Fidji Simo remains on medical leave, with plans to consolidate ChatGPT, Codex, and the AI browser into a single unified experience. Link.
Brockman wrote in a staff memo: “We’re consolidating our product efforts to execute with maximum focus toward the agentic future, to win across both consumer and enterprise.”
OpenAI told TechCrunch that although Simo remains on medical leave, she worked with Brockman on these organizational changes; Simo joined OpenAI from Instacart where she was CEO.
The consolidation reflects OpenAI’s push toward building a unified “super app” for agentic AI, with one core product team spanning consumer and enterprise.
OpenAI launched personal finance tools in preview for ChatGPT Pro subscribers, partnering with Plaid to let users connect bank accounts from over 12,000 financial institutions and ask ChatGPT questions about spending, portfolio performance, and financial planning. Link.
Users see a dashboard of portfolio performance, spending, subscriptions, and upcoming payments; the launch came one month after OpenAI acquired the team behind personal finance startup Hiro, backed by Ribbit and General Catalyst.
The feature is currently available on web and iOS for Pro subscribers only; OpenAI wants to gather feedback before expanding to Plus users.
ChatGPT currently has 800 million regular users, though only about 5% are paying subscribers; users can access the tool by selecting “Finances” in ChatGPT’s sidebar.
General Motors laid off more than 10% of its IT department, approximately 600 salaried employees, in a deliberate “skills swap” to hire workers with AI-native development, data engineering, agent and model development, and prompt engineering backgrounds. Link.
GM is still hiring for its IT department but for different skill sets; this is not a one-to-one exchange and there will likely be a net-negative job loss.
The company has been cutting white-collar employees across several departments over the past 18 months; in August 2024, GM cut approximately 1,000 software workers.
The move reflects a broader strategic transformation: rebuilding the workforce around AI skills rather than simply layering AI capabilities onto existing teams.
Anthropic updated its website to warn investors that multiple secondary platforms claiming to offer access to Anthropic shares are unauthorized, naming Open Doors Partners, Unicorns Exchange, Pachamama Capital, and others, and stating any such transactions are void. Link.
Unicorns Exchange said it received more than 50 inquiries from institutional investors wanting to buy Anthropic shares in just the past three months, with aggregate demand exceeding $1 trillion.
Both Anthropic’s preferred and common stock are subject to transfer restrictions; any sale or transfer not approved by its board of directors is considered invalid, and the company does not permit special purpose vehicles to acquire its stock.
Forge Global pushed back, claiming it was included erroneously and stated it is working with Anthropic to remove its name from the alert.


Anthropic passing OpenAI on business customer share in Ramp's data (34.4% vs 32.3%) is a more durable signal than app store rankings — procurement data reflects actual budget decisions, not curiosity. The Cerebras IPO jump to $66B on $510M revenue with 83% of that from one customer is the kind of concentration risk that gets papered over in bull markets. Those two data points together sketch a market where foundation model trust is consolidating faster than infrastructure confidence should. At theaifounder.substack.com I watch how this plays out for application-layer builders: if infra valuation concentration creates fragility, what's the worst-case scenario for a startup that built critical workflows on Cerebras inferencing specifically?