AI Labs Gear Up for AGI Amid Funding and Tensions
OpenAI closes $12.2B round at $852B valuation with $2B monthly revenue, but secondary shares stall; Anthropic secondary hits $600B as leaks and pricing hikes expose agent costs nearing human salaries.
Massive Fundraising Masks Secondary Market Fatigue
OpenAI closed a record $12.2B funding round at an $852B valuation, adding $12B from financial investors after an initial $11B from Amazon, Nvidia, and SoftBank. This included $3B from individual investors via wealth management channels, marking a first for the company. Revenue hit $2B per month, up from $1.6B late last year, growing four times faster than Google or Meta during their peaks. Yet secondary markets tell a different story: Next Round Capital reported hundreds of millions in OpenAI shares unsold despite hundreds of institutional buyers, with $2B in cash earmarked for Anthropic instead. Anthropic's primary valuation was $38B, but secondaries trade at up to $60B—seen as better risk-reward by investors like Adam Cawley of Augment Capital, betting it'll catch OpenAI. This dynamic pressures OpenAI toward an IPO, potentially in Q4 to preempt Anthropic's October target.
"We literally couldn't find anyone in our pool of hundreds of institutional investors to take these shares."
— Ken Smith, Next Round Capital, on OpenAI secondary sales.
Executive Overhauls Signal IPO and Spending Clashes
OpenAI's leadership faced multiple hits: AGI deployment CEO Fiji Simo took medical leave for a neuroimmune relapse, with Greg Brockman overseeing product, Jason Quan (CSO), Sarah Frier (CFO), and Denise Dresser (CRO) handling business ops. COO Brad Lightcap shifts to special projects like PE joint ventures; Dresser adds COO duties; CMO Katie Roush steps back for cancer recovery, interim filled by ex-Meta's Gary Briggs. CFO Sarah Frier clashes with Sam Altman on IPO timing and $60B infrastructure spend over five years, projecting $20B annual burns before decade-end profitability. Frier doubts data center needs and revenue sustainability, excluded from key investor talks—ironic given her Square IPO success. This intensifies as competition heats and IPO looms, echoing broader stakes ratcheting up.
"She is working with a founder with big ambitions who wants to push the envelope as hard as he can on spend."
— Anonymous source on Frier's challenges with Altman.
TBPN Buy Sparks Media Strategy Debates
OpenAI acquired daily tech talk show TBPN (3-hour livestreams with execs and founders) amid Fiji Simo's 'no side quests' mandate—ironically pushed by her. Reactions split: Wharton prof Paul Ner called it nonsensical; NYT's Mike Isaac saw tech-media frustration; Slow Ventures' Jack Reichs viewed it as premium for relationships, like ML researcher paydays. TBPN retains strict editorial independence, per founders John and Jordy, but critics like Simon Smith argue it undermines OpenAI's productivity focus or becomes a true side quest. Not a distribution play (Sam Altman's X reach dwarfs it), nor likely AI-training data goldmine per Robert Scoble. Core motive: Counter controversy stigma by capturing TBPN's positive insider vibe, though it risks alienating competitors and misses broader public narrative battles.
Anthropic's Turbulence: Leaks, Limits, and Pricing Shifts
Anthropic's Claude Code update leaked 512k lines of code, revealing unreleased features like always-on agent 'Chyros' (background work, dream mode for memory consolidation, proactive mode), Tamagotchi-like 'buddies' for Twitter buzz, five compaction strategies, dozens of tools, caching for sub-agents, and configurable prompts. Human error blamed; lawyer retracted 8,000+ erroneous GitHub takedowns. Usage limits frustrated Pro/Max users (caps in minutes), chalked to tighter peak limits and million-token sessions by co-dev Lydia Hi—advising Sonnet, lower effort, fresh sessions. Friday's hike blocks subscription use for third-party tools like OpenClaw, forcing pay-per-token API. Creator Peter Steinberger delayed it a week; signals end of subsidies.
"One thing is clear from reading the code: harness engineering is hard and deeply non-trivial."
— Euchen Jin on Claude Code complexity.
Agent Economics Demand Real Math, Not Hype
Anthropic's moves spotlight rising costs: Agents on bleeding-edge chips (depreciating in 3 years, nuclear-powered datacenters) aren't 'pennies.' Daniel Jeff warns the 'agent economy' mirrors full-time salaries—transformers guzzle memory, best models eat hardware fastest. Older tasks cheapen for on-device, but Mac Studio clusters aren't free. Smashes 'AI replaces all jobs cheap' myth; true intelligence costs like hiring teams, often pricier/less reliable than humans. Subsidy era ends fast in AI age, reshaping pricing, jobs debates, and feasibility.
"The agent economy is not cheap... intelligence going up into the right keeps eating the bleeding edge of the best chips... as fast as we can make them."
— Daniel Jeff on true costs.
Model Releases Highlight Open vs. Proprietary Split
Google's Gemma 4 open-source family (2B/4B for edge, 26B MoE, 31B dense) hits #3 on Arena leaderboard, frontier-per-parameter efficiency on Gemini architecture. Strong coding/agentic, 256k context, 140 languages, runs on laptops—swap into Cursor, Hermes, OpenClaw. Alibaba's proprietary push: Qwen 3.6 Plus (multimodal, 1M context, 1/8 Opus cost, near Swebench parity) after two others in three days. CEO Eddie Wu takes AI lead post-researcher exits, prioritizing revenue.
Datacenter Risks from Geopolitical Shocks
Unmentioned in depth but thematic: Energy shocks expose vulnerabilities as $60B infra bets collide with global tensions, amplifying storm risks.
Key Takeaways
- Track secondary valuations over primaries: Anthropic's $600B implied beats OpenAI's $852B for risk-reward in pre-IPO bets.
- Brace for agent pricing reality: Expect costs akin to salaries, not pennies—factor into product economics now.
- Leadership friction accelerates with scale: OpenAI's IPO push vs. spend skepticism shows founder-CFO tensions as table stakes.
- Media acquisitions like TBPN prioritize narrative control over tech insiders, but public perception remains the real battleground.
- Open models like Gemma 4 enable local frontier runs: Test 31B on laptops for coding/agent workflows to cut API reliance.
- Leaks reveal harness complexity: Build wrappers post-distribution like Cursor, not from scratch.
- End of subsidies forces API shifts: Migrate third-party tools (e.g., OpenClaw) to pay-per-token before lockouts.
- Revenue growth hides burns: OpenAI's 4x historical pace won't cover $20B/year without profitability paths.