Bitcoin just had its worst week in months. Price cratered below $60,000 on Friday, hit a low of $59,227 overnight, and triggered $1.6 billion in liquidations. The trigger was a hot May jobs report that revived rate-hike fears and sank the Nasdaq 100 about 5%. Meanwhile, Michael Saylor's Strategy quietly disclosed it sold 32 BTC at the end of May, which traders pounced on, even though the actual selling pressure came from somewhere much bigger. AI was a big part of that somewhere. Mati Greenspan, Saylor, and Jameson Lopp all pointed at the AI capital boom as draining money out of bitcoin. And the AI side keeps delivering. xAI is shipping Grok Build to all developers, BMW just hit 30,000 cars assembled with help from Figure humanoids, and the FDIC dropped its first real rulebook for stablecoin issuers under the GENIUS Act. Four stories worth your time.
Let's start with xAI, because Elon's AI company is doing something genuinely interesting and it's getting almost no coverage compared to OpenAI and Anthropic. Three things happened in the last few days. First, xAI deepened its partnership with Cloudflare. Every Grok model, including Grok 4.3 and the new Grok Build 0.1, is now fully integrated into Cloudflare's AI Gateway. That means developers get unified logging, caching, rate limiting, spend controls, and one consolidated invoice across providers. Cloudflare is quietly turning itself into the control plane for AI API traffic, and xAI just made it the easiest place to deploy Grok at scale. Second, Grok Build 0.1 is now in public beta for any developer, no SuperGrok subscription required. This is xAI's answer to Claude Code. It runs at 100-plus tokens per second, costs $1 per million input tokens and $2 per million output, supports a 256k context window, and can run up to 8 agents in parallel using isolated Git worktrees. It also has native Model Context Protocol support, so you can plug in your own internal knowledge bases and tools. Claude Code is still better for deep engineering work, but for parallel-heavy migrations and high-volume API workloads, Grok Build is now a real option. Third, and this is the one that should make every coding assistant company nervous. There's a report that Grok V9-Medium, xAI's next model at 1.5 trillion parameters, is being trained on real developer workflows from Cursor. Not public GitHub code. Actual private codebases, real multi-file edits, real interaction patterns. And xAI reportedly holds a $60 billion option to acquire Anysphere, Cursor's parent company. If they exercise that option, they own the largest AI code editor and lock competitors out of the data pipeline. The parameter count isn't the story. The training data moat is the story. Whoever owns the workflow owns the model.
Now to the other AI story that's actually showing up in physical results. Humanoid robots are crossing the line from demo videos into real production environments. BMW just shared numbers from its 11-month pilot with Figure 02 robots in Spartanburg, South Carolina. Two robots logged about 1,250 runtime hours, handled over 90,000 parts, and contributed to assembly of more than 30,000 X3 vehicles. That's not a stunt. That's industrial deployment with real throughput. BMW is now expanding the program to Leipzig with Hexagon Robotics' AEON humanoids for high-voltage battery assembly, with full pilots planned this summer. The analyst framing is what matters here. For this to work economically, payback needs to be under 3 years, with uptime and task success above 95%. BMW is treating this as a manufacturing tool, not magic. Targeted pilots, continuous measurement, share the floor with humans. Figure also just demonstrated a 24-hour continuous package-sorting run with 3 humanoids. Nvidia partnered with Unitree to bundle the H2 humanoid body with Jetson Thor compute as a research platform called Isaac GR00T, shipping to Stanford and ETH Zurich. 1X started full-scale production of its NEO humanoid in California, marketed as quiet enough for the home. And Tesla's Optimus Gen 2 is being reviewed for India, with pricing estimated between 15 and 30 lakh rupees, roughly $18,000 to $36,000. The Optimus narrative is shifting in a way founders should notice. A $30,000 price point reframes humanoids as a capital investment, not a sci-fi gadget. The real value isn't a multitask robot. It's a software-enabled platform that can do one narrow task safely and repeatedly at scale. Hardware becomes the delivery shell. The recurring revenue is software, fleet management, safety updates, and task libraries. Locomotion is mostly solved. General-purpose manipulation in messy environments is the remaining wall. Warehouses, factories, and back-of-house retail get this first. Homes come much later.
Back to bitcoin. The price action this week was ugly. BTC fell below $60,000 on Friday for the first time since October 2024, briefly touched $59,227, and recovered above $61,000 by the weekend. About $1.6 billion in liquidations. The Nasdaq 100 dropped about 5%. Memecoins got crushed, with Dogecoin and Shiba Inu down 9%. XRP fell toward $1.10. Ether broke down to a 13-month low under $1,600, helped along by a serious Zcash vulnerability that an AI model uncovered, which by the way is a story worth watching, because the same researcher is now pointing his AI audit at Monero and other privacy coins. The proximate cause was a hot May payrolls report that revived rate-hike fears. The deeper cause, according to Greenspan, Saylor, and Lopp, is the AI capital boom pulling money out of bitcoin and into AI infrastructure plays. Traders also blamed Saylor after Strategy disclosed selling 32 BTC for $2.5 million at an average $77,135 per coin between May 26 and May 31. 32 coins is nothing. It does not move the market. But it broke the never-sell narrative, and that was the real damage. Saylor's response was an essay titled "disciplined expansion" arguing bitcoin needs to grow through banks, credit, and securities while preserving the base layer. There's also a quieter theory floating around. With the SpaceX IPO opening for trading on June 12, and Kraken now offering tokenized SpaceX access through xStocks, retail might be rotating out of crypto into the IPO. CoinDesk looked at exchange flows and stablecoin movements and didn't find clear evidence of a mass exit yet. The real reporting comes in July. Funding rates flipped negative, $2.6 billion in short leverage piled up, and analysts are now watching for a potential short squeeze. Seller exhaustion may be setting in. Or it may not. The diehards are buying the dip. Jack Mallers said as much. Everyone else is waiting for the Coinbase premium to flip positive before calling a bottom.
Last topic. The GENIUS Act is no longer abstract. The FDIC just dropped two notices of proposed rulemaking that turn permitted payment stablecoin issuers into something that looks a lot like banks. Here's what's in the proposals. FDIC-supervised stablecoin issuers will need full Bank Secrecy Act and anti-money laundering programs, written customer identification programs, beneficial ownership procedures, suspicious activity reporting, and OFAC sanctions compliance. They'll face routine FDIC examinations and enforcement actions. Civil penalties can hit $100,000 per day for material violations. FinCEN and OFAC already released a joint proposal in April treating these issuers as financial institutions for AML and sanctions purposes, including the ability to block impermissible transactions and comply with lawful orders. Public comments closed June 9. This is not light-touch. The U.S. is bringing stablecoin issuance fully inside the banking regulatory perimeter. Two implications. First, Circle and USDC look well positioned. Circle's reserves are already in short-term Treasuries and cash equivalents, custodied at BNY Mellon and BlackRock. Monthly disclosures, 1:1 backing. That's basically the GENIUS Act template. Holders even get priority in bankruptcy under the new framework. USDC traded at $0.9996 on June 3, slightly below peg, after $500 million was minted on Solana in 48 hours, but volume was 3.4 times the 30-day average, suggesting institutional repositioning rather than stress. Second, Tether is going a different direction. Tether teamed up with Canaan and ACME Swisstech to build a modular Bitcoin mining system with swappable hash board modules, immersion cooling, and Tether's own control software. Paolo Ardoino is doubling down on becoming a major bitcoin miner, not a U.S.-regulated payments company. Visa, meanwhile, is testing private stablecoin settlement on Canton with Brale, exploring whether institutions can use blockchain without leaking transaction data. The bigger picture. Stablecoins are splitting into two camps. The U.S.-regulated bank-adjacent path that Circle is sprinting down. And the offshore, bitcoin-aligned, infrastructure-heavy path that Tether is building. Both can win. They're not competing for the same customer anymore.
One observation to end on. The same week bitcoin crashed because AI is sucking up all the capital, AI also exposed a 4-year-old cryptographic flaw in a major privacy coin and is now training on real developer workflows to build the next coding assistant. The technology eating bitcoin's lunch is also auditing its code. Decide for yourself whether that's a threat or the best possible adversary.