Money moved all weekend. Alibaba pulled in $10.2 billion in the biggest share sale Hong Kong has ever seen, all of it earmarked for AI. Hugging Face, the default home of open weights, is reportedly weighing a sale north of $13 billion. And Nvidia is in talks to back Perplexity at a $30 billion-plus valuation. In the middle of all that, OpenAI quietly cut GPT-5.6 Sol’s API price by over 20% for three months. The funding machine and the price war are running in parallel. That’s what a real market looks like.

The Money Is Chasing AI Again

Alibaba: $10.2 billion for AI, and a 10% share drop

Alibaba priced a $10.2 billion share placement in Hong Kong over the weekend, the largest primary follow-on offering ever by a Hong Kong-listed company, per Bloomberg and CNBC. All net proceeds go to AI infrastructure, and the deal was oversubscribed. The market’s reaction tells you everything: shares plunged 10% on the dilution. Wall Street wants Alibaba to spend on AI, then punishes the company for doing exactly that. That tension is the whole AI capex story in one tick.

Hugging Face: $13 billion sale talk

Business Insider and Bloomberg report Hugging Face is exploring a sale that could value it at $13 billion or more, roughly triple its 2023 round. The platform hosts millions of models, runs the Inference Providers network, and has become the default distribution layer for open weights. A sale at that number would be a signal even the neutral host sees an exit in the current market. Someone is doing math on the hub, not the spokes.

Nvidia circles Perplexity at $30 billion

Reuters, citing The Information, reports Nvidia is in talks to invest in Perplexity as part of an equity round valuing the search startup above $30 billion, up more than 50% from its ~$20 billion round two months ago. Perplexity’s annualized revenue reportedly tops $750 million. Nvidia’s playbook here is familiar: take an equity stake in whoever needs the most GPUs, from OpenAI to Anthropic to xAI. It’s not the founder. It’s the landlord.

OpenAI’s price cut is real, but it’s a three-month smell

OpenAI confirmed it’s dropping GPT-5.6 Sol API and credit pricing by over 20% for the next three months, starting August 22. Standard short-context pricing drops to $4 per million input tokens and $20 per million output tokens, and Codex credits and ChatGPT Work get the same cut. Luna already fell 80% to $0.20 earlier. The company frames it as pushing the price-performance frontier as efficiency improves.

Here’s the honest read: OpenAI is conceding on price, not on frontier. DeepSeek V4 Flash undercuts Sol by orders of magnitude, Qwen3.8-Max is open weight, and the three-month window says more than the number does. A timed discount is a customer-retention move, not a strategy. When November rolls around, either the next model justifies a price rebound or the cut becomes permanent. I know which one I’d bet on.

The chips and talent arms race goes vertical

Anthropic hires the Google TPU founder

Anthropic hired Amir Salek, founder and former head of Google’s TPU program, for its compute team as it builds custom silicon, Bloomberg reported August 21. It’s Anthropic’s second major chip hire after pulling an OpenAI engineer in June. Custom silicon is expensive, slow, and politically loaded. Doing it anyway tells you how bad the margin squeeze feels at frontier-inference scale.

Luke Metz leaves OpenAI for Meta

Axios confirmed August 24 that AI researcher Luke Metz, who returned to OpenAI from Mira Murati’s Thinking Machines Lab in January, is joining Meta’s Superintelligence Labs and will report to chief AI officer Alexandr Wang. Metz co-developed DCGAN and worked on o1. Talent moves between frontier labs like transfers in football now, and every move is a strategic signal about whose bet is winning.

The data center backlash gets a governor

Texas Governor Greg Abbott told Newsweek that AI data centers “dug their own grave” with local communities, and confirmed Texas has paused approvals for new data centers while statewide standards are drafted. Axios reported the same, framing it as data-center spending now large enough to matter for the whole economy. This is the first big governor coming out explicitly against the infrastructure land grab, and it follows months of power, water, and noise complaints. Every hyperscaler with Texas ground in play just got a new risk item.

Quick Hits

– Prime Intellect – Published results of 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun (August 19); Fable 5 topped the field at 82% of the human record. The real finding was how wide the spread between the best and worst autonomous agents is.

– Humanoid robots – A robot called Lightning ran 100 meters in 9.39 seconds at the World Humanoid Robot Games in Beijing (August 22), faster than Bolt’s 9.58 human record, in a preliminary heat. Reuters had it first.

– Grok’s web chat – Adversa AI demonstrated a zero-click attack technique, Cryptographic Context Injection, that can exfiltrate chat history from Grok via AES-encrypted payloads. xAI had not patched it as of August 23.

– DeepMind – “From Atari to EVE Online” (August 21) recaps 15 years of game AI research, from DQN to AlphaStar, and sets up the next chapter with game developer partners.

Rundown for August 24. Sources: Yellow, OpenAI, CNBC, Bloomberg, Reuters, Axios, Newsweek, The Information, CBS News, DeepMind, Prime Intellect.