The Calm Before the AGI Storm

April 6, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
April 06, 2026

There was no single massive story last week — no new frontier model, no overnight paradigm shift. And yet when you take the sum total of what happened, there's a very distinct picture emerging. The big labs are all jostling and positioning for a fast-moving future, and there's an electric charge in the air — the calm before the AGI storm.

OPENAI

$122B Round, $852B Valuation — and Secondary Market Trouble

OpenAI began the week on a high, closing a record-breaking $122B funding round at an $852B valuation — the largest venture round ever. The additional $12B (above the previously announced $110B from Amazon, NVIDIA, and SoftBank) came largely from financial investors, including $3B from individual investors through wealth management channels for the first time. OpenAI also disclosed they're now generating $2B in revenue per month, up from $1.6B at year end, and claim they're growing four times faster than the companies that defined the internet and mobile eras. But the picture got complicated quickly: Bloomberg reported that OpenAI stock is struggling to find buyers in secondary markets, with hundreds of millions worth of shares going unwanted while buyers are reportedly sitting on $2B in cash ready to deploy into Anthropic instead. As Adam Crowley of Augment Capital put it: "It's just better risk-reward right now. People are betting that Anthropic's valuation will catch up with OpenAI's. But if you buy OpenAI shares, it's less clear what the return will be in the near term."

C-Suite Reshuffled at the Worst Possible Time

Fidji Simo, CEO of AGI Deployment, announced she's taking several weeks of medical leave after pushing herself too hard — deferring tests and new treatments for a chronic neuro-immune condition since joining the company. President Greg Brockman will run product in her absence. There were also several other departures: Brad Lightcap steps down as COO to lead OpenAI's private equity joint venture push; Denise Dresser takes on the COO role on top of her CRO duties; and CMO Katie Rouch has stepped down to focus on her cancer recovery, with former Meta CMO Gary Briggs filling in. In isolation, none of this reads as the ship-jumping kind of executive shakeup. It feels more like terrible timing than Machiavellian psychodrama. But an executive reshuffle is still the last thing OpenAI needs heading into an IPO year with Anthropic breathing down their neck.

Altman and CFO Are at Odds Over the IPO

Over the weekend, The Information reported that Altman and CFO Sarah Friar are in disagreement over IPO timing and spending. Altman wants to go public in Q4 — possibly even before Anthropic, who are targeting October. Friar reportedly doesn't think OpenAI will be ready this year, and has expressed skepticism about the data center spending commitments and whether revenue growth can support them — concerns that are notable given Altman has committed to $600B in infrastructure spending over five years and forecasts burning $200B before turning profitable. Sources say Friar has even been excluded from some capex discussions with investors, which were described as noticeably awkward given she was present before. She was explicitly brought in to have exactly these kinds of tough conversations — she previously shepherded Square to a successful IPO. One source summarized it: "She's working for a founder with big ambitions who wants to push the envelope as hard as he can on spend."

OpenAI Buys TBPN

The other major OpenAI story was their acquisition of tech talk show TBPN — a daily three-hour video podcast featuring tech executives and founders — reportedly pushed for by Simo herself, which made it awkward given her simultaneous crusade against side quests. Reactions were all over the map. Wharton's Paul Neri called it baffling from an M&A perspective. Mike Isaac argued it was the clearest proof yet of CEO frustration with mainstream media's AI coverage, framing it as a marketing expense that sets a price floor for every other tech-adjacent show. Slow Ventures' Jack Rains pushed back on that: Sam Altman's Twitter has bigger distribution than any tech media platform — this isn't about buying reach. My read is simpler: for more than a year OpenAI has felt the sting of being the company most associated with technology that is deeply publicly controversial, and TBPN has good media juju. The logic is buy them and see if you can get some of that lightning in a bottle. The tension is real though: either the show maximally supports OpenAI's focus, in which case editorial independence is a fiction, or it has genuine editorial independence, in which case it's a side quest.

ANTHROPIC

Source Code Leak

Anthropic had one hell of a week. An update to Claude Code accidentally included 512,000 lines of source code. Anthropic yanked it quickly but not before the internet had it — then spent the following day issuing over 8,000 GitHub DMCA takedowns, most of which were targeting publicly released forks rather than the leaked proprietary code. Boris Cherny later apologized and retracted all but one. The leak revealed some interesting unreleased features: an always-on background agent called Kairos (complete with "dream mode" for autonomous memory consolidation and a "proactive" mode for taking initiative without instructions), and a Tamagotchi-style virtual pet feature called Buddies, whose source code revealed Anthropic hoped to generate "sustained Twitter buzz" from it. The deeper takeaway from people who read the code: Claude Code is far more complex than it looks — five different context compaction strategies, caching optimizations for subagents, dozens of tools. As Yuchen Jin put it, "harness engineering is hard and deeply non-trivial."

Usage Limit Scandal and an OpenClaw Crackdown

Separately, usage limits became a scandal in their own right, with users reporting they were burning through their allocations in hours or even after a single Opus prompt. Claude Code developer Lydia Hallie's official response — essentially "peak-hour limits are tighter and million-token context sessions got bigger, you're not being overcharged" — landed badly, with many hearing it as "you're holding it wrong." And then on Friday, Anthropic announced that subscribers can no longer use their subscription to power third-party tools like OpenClaw — requiring them to pay via the API instead. OpenClaw creator Peter Steinberger said he tried to talk Anthropic out of it and only managed to delay by a week. Daniel Jeffries had the most salient take: the subsidy era is ending. Running super-intelligent agents around the clock on the most expensive chips ever made was never going to be cheap, and the agent economy will be priced more like a human salary than a SaaS subscription.

MODELS AND CHINA

Gemma 4

Google made a significant move in open source with the release of Gemma 4 — the first Western open source model competing at this level in years. The 31B dense model ranks third on the Arena AI text leaderboard for open source, and the lineup runs from 2B edge models all the way up to a 26B mixture of experts. The models are optimized for coding and agentic performance, built on the same architecture as Gemini, and can run on a laptop. As Greg Isenberg put it: a few months ago running something this capable locally meant serious hardware tradeoffs. Now it works offline on your phone, speaks 140 languages, has a 256K context window, and costs nothing — and you can swap it in as your model in Claude Code, Cursor, or OpenClaw right now.

Alibaba's Pivot, and Deepseek V4 on the Horizon

Meanwhile in China, Alibaba is moving in the opposite direction from open source. They released three proprietary models in three days culminating in Qwen-3.6 Plus — which lags Opus 4.5 by a few points on SWE-bench but comes in at roughly one eighth the cost and achieved the first trillion tokens served on release day. CEO Eddie Wu has personally taken over the AI division with a revenue maximization mandate, and early returns validate it. And all eyes are on Deepseek V4, now expected in the next few weeks — and set to be a watershed for Chinese semiconductors. Orders for Huawei chips are flooding in from every Chinese tech giant, with Deepseek having spent months optimizing the model specifically to run on Huawei hardware. If it works at scale, this could mark the beginning of the end of China's dependence on NVIDIA.

Microsoft Targets Superintelligence by 2027

Microsoft released three new models for transcription, voice, and image generation — nothing externally notable, but a signal that they're back in the model training game after M-AI-1 preview went dark last August. These models will be deployed under the hood in Teams and other products as a cost-cutting measure. The bigger story is Mustafa Suleiman's ambition: in an interview with Bloomberg he said "we must deliver the absolute frontier — certainly by 2027, the objective is to really get to state-of-the-art," backed by a GB200 Blackwell training cluster coming online in October. Separately, Copilot sales are back on track — commercial CEO Judson Althoff told staff they'd hit their "big audacious goals" for Q1, with Microsoft now pitching Copilot as a unified access point for all the best models including Anthropic's.

DATA CENTERS

Data Centers Under Pressure — Energy, Supply Chains, and Iran

The Iranian Revolutionary Guard declared 18 US tech companies — including NVIDIA, Apple, Microsoft, and Google — "legitimate targets" for retaliation, explicitly linking them to AI-enhanced targeting in the war. With three Amazon data centers in Bahrain and the UAE already hit by drones, this makes further Middle East construction look untenable. The energy shock is rippling through Asia too, where $800B in data center projects are planned through the decade and financing is suddenly a much harder conversation. In the US the bottleneck is different: more than half of planned data centers are expected to face delays not because of money, but because transformers, switchgear, and batteries are in short supply. As Crusoe's Andrew Likens put it: "If one piece of your supply chain is delayed, then your whole project can't deliver. It's a pretty wild puzzle at the moment."

WHAT COMES NEXT

Spud and a New Social Contract

Three image models — codenamed Masking Tape, Gaffer Tape, and Packing Tape — appeared on Arena AI over the weekend and turned heads, with some speculating they're the first look at Spud, which would be OpenAI's first LLM with native multimodal training. Word from inside the labs is that the next generation of models represents a genuine major shift. And then, just as recording began, OpenAI dropped what amounts to a new social contract thought-starter. The first line says it all: "As we move towards superintelligence, incremental policy updates won't be enough." That'll be a big part of tomorrow's show — but it's also the clearest indication yet that the calm before the AGI storm is not going to last long.