Friday, May 29th. Here's what's moving. Elon Musk says Grok V9-Medium has finished training, 1.5 trillion parameters, public release in two to three weeks, with coding as the explicit target. Figure AI signed its first real commercial deployment, putting humanoid robots into a Catalyst Brands distribution center in Reno. Bitcoin is pinned below $74,000 with spot ETFs bleeding for a record 9 straight days, $2.8 billion gone. Paxos became the first blockchain-native clearing agency approved by the SEC, sitting alongside DTCC. And six federal agencies are racing to write GENIUS Act rules before the July deadline, while the Federal Reserve has somehow not published anything yet. Let's get into it.
xAI says Grok V9-Medium is done training. 1.5 trillion parameters, roughly three times the size of V8-small, optimized for Nvidia Blackwell GPUs, and pointed squarely at the coding market. Public release, mid-June.
The interesting part isn't the parameter count. It's the training data. xAI leaned heavily on Cursor data, real developer workflows from the editor used inside OpenAI, Stripe, and basically every serious AI-native engineering team. That's a different bet than scraping public GitHub. Public GitHub is a graveyard of abandoned side projects. Cursor data is what working engineers actually do when they're shipping.
The gap they need to close is real. On SWE-bench Verified, Grok v8-small sits around 75%. Claude Opus 4.7 is at 87.6%. GPT-5.5 is around 88.7%. And in enterprise AI adoption, Grok is at about 6%, versus 55% for OpenAI, 47% for Anthropic, 39% for Google. Grok is the fourth horse in a four-horse race.
There's also Grok Build, xAI's first coding agent, already in early beta, aimed directly at Claude Code. Current context window is 256K tokens, which is a real bottleneck for large codebases, and V9-Medium is supposed to fix that.
Two things to watch. One, does Cursor-trained data actually translate into better real-world coding, or just better benchmark scores? Those are not the same thing. Two, xAI says it will open-source the older 0.5T V8 model later this year. That's a meaningful contribution if it actually ships. Meta used to play that role and has gone quiet. If Musk wants developer mindshare, open-sourcing a half-trillion-parameter model is a faster route than any benchmark.
Mid-June is the deadline he set himself. We'll know soon enough whether the hype matches the weights.
Figure AI signed a real commercial deal with Catalyst Brands, the company formed from the JCPenney and SPARC Group merger that runs Aéropostale, Brooks Brothers, and a few hundred storefronts. The Figure 03 humanoids are going into Catalyst's Reno, Nevada distribution center, working alongside the Joey Pouch sorting system Catalyst already spent $40 million upgrading.
The pre-deployment numbers are worth taking seriously. Figure ran a 200-hour endurance test. In a separate 72-hour run, the robots sorted 88,000 packages. In another, they processed about 250,000 packages over 52 hours with no hardware failures. That's not a demo reel. That's a robot doing the actual job.
The AI side is also worth noting. Figure's running its Helix-02 model locally on the robot, not streaming inference from a cloud GPU somewhere. That matters for warehouse deployments where you can't have your sorting line stop because AWS had a bad afternoon.
Meanwhile, Tesla. On May 21st, Optimus was filmed handing a water bottle to a person in an unscripted setting. Sounds trivial. It isn't. Grasping a cylindrical object, locating a hand it hasn't seen before, releasing without dropping, in an unmapped environment, that's generalization. Tesla is converting the Fremont Model S and X line for Optimus production, targeting late July or August.
The broader picture. Analysts project around 50,000 humanoids produced annually by 2028. Tiny compared to global manufacturing labor, but this is the inflection point where they stop being a YouTube demo and start being a line item on a logistics budget. Figure is building toward 100,000 units over four years at its BotQ facility.
The interesting competitive question isn't Figure versus Tesla. It's who locks in the long-term retail and logistics contracts first. Catalyst is the first publicly announced humanoid deployment in U.S. retail logistics. The second one will come faster.
Bitcoin is having a bad week, and the data underneath is uglier than the price.
Price action first. BTC is pinned below $74,000, dipping toward $72,000, near April lows. Meanwhile the S&P 500 and Nasdaq futures are touching all-time highs. Oil cracked on a tentative U.S.-Iran ceasefire extension. Risk assets are partying. Bitcoin isn't invited. Market cap dropped below $1.5 trillion, knocking BTC out of the global top 10 assets, passed by AI stocks and precious metals.
U.S. spot Bitcoin ETFs have now seen 9 straight days of outflows, $2.8 billion withdrawn. That's the longest losing streak since the ETFs launched in January 2024.
Here's the part that should make holders pay attention. CryptoQuant says long-term holder supply just hit a record high. Normally that's a bullish signal, diamond hands accumulating. CryptoQuant's read is the opposite. It's not strong conviction holding the line. It's the absence of new buyers. Coins are stuck with existing holders because nobody new is showing up to take them. Major investor cohorts have stopped accumulating. That pattern has historically preceded sustained weakness.
There's a $9 billion options expiry today, with bears holding the upper hand. Strategy is under capital structure pressure, with Arca's Jeff Dorman calling the situation out of hand over $15 billion in preferred stock obligations, and the CEO openly floating possible Bitcoin sales. A French company, Sequans, just abandoned its Bitcoin treasury strategy entirely and is liquidating.
Some counterweight. Texas is moving its $10 million Strategic Bitcoin Reserve out of BlackRock's IBIT ETF into direct custody, hiring an actual crypto custodian. Fidelity Digital Assets is publishing notes about growing evidence of a shift away from dollar-based settlement, with nation-states leaning on Bitcoin and gold. Retail is buying the dip. But spot demand is weak, ETF demand is weak, and the next catalyst, according to most analysts, isn't geopolitical. It's regulatory.
On the regulatory front, two stories that fit together.
First, GENIUS Act implementation. The law passed last July, final rules due July 18th. Six federal agencies have moved. The OCC published a 376-page proposed rule back in February. Treasury, FDIC, FinCEN and OFAC jointly, and the NCUA have all issued notices of proposed rulemaking. Comment windows are closing through June and July. The Federal Reserve has published nothing. Zero. Entities under Fed supervision will have dramatically less time to prepare than entities under OCC or FDIC, which is going to create a real compliance scramble.
The substance is bank-like. Permitted issuers must hold 1-to-1 reserves in cash, Fed balances, insured deposits, short Treasuries under 93 days, or qualifying money market funds. Monthly reserve disclosures, annual audits, full AML programs, OFAC sanctions screening, ability to block and freeze illicit transactions. Stablecoin issuers are now financial institutions under the Bank Secrecy Act.
Which brings us to Tether. Tether's playing a clever game. They launched USAT, a U.S.-domiciled, GENIUS-compliant stablecoin, custodied by Cantor Fitzgerald, vetted by Deloitte, sitting inside Anchorage Digital Bank. USAT grew over 500% in a month, now above $140 million market cap, though still well behind USDC, PYUSD, and RLUSD.
Meanwhile USDT stays offshore. Same dollar peg, but reserves that include gold and Bitcoin, no monthly attestations to U.S. standards, no GENIUS Act compliance. The interpretation gaining traction is that this split is the entire point. USAT exists so USDT never has to comply. By mid-2028, U.S. exchanges will likely have to delist non-GENIUS stablecoins. USAT picks up the U.S. market. USDT keeps the offshore market, with its high-yield reserve mix completely outside U.S. oversight.
It's a regulatory firewall. The compliant entity absorbs the scrutiny. The lucrative entity keeps doing what it was doing. Whether regulators tolerate that arrangement long-term is the question. Treating USDT activity on U.S. intermediaries as a sanctions exposure problem is the obvious lever, and the FinCEN-OFAC joint rule already gestures in that direction.
One related note. SEC Commissioner Hester Peirce came out defending crypto privacy tools against the surveillance push, arguing privacy-enhancing tech actually strengthens investor protection. Worth watching as the AML rules get finalized.
One prediction. If Bitcoin's long-term holder supply is at a record high not because of conviction but because no new buyers exist, the next leg down isn't going to be triggered by macro. It'll be triggered by one large holder deciding they need liquidity more than they need their thesis.