Monday, August 10th. Bitcoin is holding above $65,000 heading into a big week for macro data, with July CPI due Wednesday and the Treasury auctioning $125 billion in notes across three days. A weak jobs report Friday, negative 23,000 payrolls, killed off September Fed hike bets and gave risk assets a lift. Nvidia's Vera Rubin platform is now in full production and shipping to hyperscalers, though the company is quietly trimming HBM memory on Rubin Ultra to deal with a global memory squeeze. Strategy sold 1,690 bitcoin to fund a $108 million buyback of its STRC preferred shares. And humanoid robots are showing up on factory floors from Kyoto to Guangzhou faster than most people expected. Let's get into it.
Bitcoin is in a strange spot right now. Price steady in the mid $65,000s, briefly touching $65,245, but under the surface things are shifting fast. Friday's jobs report was ugly. Payrolls fell 23,000 in July, the first negative print since 2020. Revisions wiped out another 103,000 jobs from May and June. Wage growth cooled to 3.2%. Participation dropped to a post-pandemic low. That was enough to yank September rate hike odds on CME FedWatch from around 55% down to roughly 40%. Ten-year yields fell to about 4.65%. Dollar softened. Gold pushed above $4,400. Classic risk-on setup, and bitcoin rode along.
But here's what makes this week interesting. The Treasury is auctioning $125 billion between Tuesday and Thursday. $58 billion in 3-years, $42 billion in 10-years, $25 billion in 30-years. About $28.7 billion of that is new cash. And it lands right on top of July CPI Wednesday and PPI Thursday. If demand at those auctions is soft, yields spike, and the 30-year is already flirting with 5.2%. If CPI runs hot, hike bets come back and this rally stalls.
Derivatives are telling a mixed story. Bitcoin's BVIV volatility index just hit its lowest level since 2025. Option demand collapsed. But downside protection still commands a premium, so traders aren't exactly relaxed. Meanwhile, CME leveraged funds just flipped net long, abandoning the structural short basis trade that's dominated positioning for years. That's a real shift. When hedge funds stop shorting futures and start actually betting on upside, it usually means the easy carry trade broke and they're repositioning for direction.
One more thing worth flagging. Strategy sold 1,690 bitcoin over the weekend, raising $108.6 million, and used it to buy back STRC preferred shares. Bitcoin holdings now sit at 840,447. Cash reserves climbed to $4.65 billion. That's a treasury company selling BTC to prop up its own capital structure, which is a story we're going to see more of as these vehicles mature.
Nvidia's next-generation platform, Vera Rubin, is officially shipping. Full production, deployed to OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta, Dell, and others. The NVL72 rack packs 72 Rubin GPUs and 36 Vera CPUs, fully liquid-cooled, connected by NVLink 6. Nvidia claims up to 10 times higher inference throughput per watt at one-tenth the cost per token versus Blackwell. Installation drops from 2 hours to 5 minutes thanks to cable-free modular trays. This is a real generational leap, and Microsoft is putting these into its Fairwater AI data centers.
But here's the twist. Nvidia is quietly trimming the memory on Rubin Ultra. Jensen Huang previously touted 1 terabyte of HBM4E per GPU. Test samples are now showing 192 to 256 gigabytes. Some configurations may even fall back to older HBM4. The reason is a genuine global shortage of high-bandwidth memory. SK hynix, Samsung, Micron, none of them can produce 12-high or 16-high HBM4E stacks fast enough. So Nvidia is stepping down from 16-high to 12-high, offering 8-high variants, and even cutting LPDDR5X on the Vera CPU side by about 50%.
Nvidia is framing this as customer choice. Configurable SKUs. Modular SOCAMM architecture that scales from 96 gigabyte modules up to 1.5 terabytes. They're also pushing storage-class memory alternatives like BlueField-4 STX with CXM context memory for long-context inference. That's a legitimate architectural pivot, not just spin.
Still, the underlying reality is that memory supply, not compute, is now the binding constraint on AI hardware. TrendForce says the shortage will get worse into next year. AMD's MI400 is facing the same despec pressure. Analysts expect dramatic year-over-year gains in HBM operating profits at Samsung and SK hynix. The pricing power has shifted from GPU makers to memory makers, at least for this cycle. If you're wondering why every hyperscaler is signing 5-year memory supply deals right now, that's the answer.
Humanoid robots are moving from demo videos to actual production lines faster than most forecasts suggested. Three stories from the past week make that clear.
Mitsubishi Motors is converting an unused engine production line at its Kyoto plant to mass-produce humanoid robots. Target output: up to 1,000 units per month. Start of production: first half of 2027. The robots come from Highlanders, a University of Tokyo startup. They're about 1.75 meters tall, have 19 degrees of freedom, use AI perception and natural language interaction. Mitsubishi will deploy them first in its own factories to deal with Japan's labor shortage, then likely sell to other manufacturers. When a car company converts an engine plant to build humanoids, that's a signal about where they think the durable demand is.
In Europe, a startup called Nucleus deployed humanoid robots inside a major German factory within 90 days of announcement. They're using off-the-shelf Unitree G1 hardware with custom end effectors, focusing entirely on the software and workflow layer. The model is supervised autonomy. Humans guide the tricky parts, robots handle the repetitive work, and every hour of operation generates training data. Clients pay hourly for actual work performed. It's basically Waymo's playbook applied to factories, treating real-world operating data as the moat.
Meanwhile, Avatar Robotics closed a $6.5 million seed led by AlleyCorp for essentially the same thesis in the US. Semi-humanoid robots with teleoperation, human-in-the-loop, targeting logistics and warehousing. They've already processed over 900,000 products in live customer facilities and are expanding with a multi-billion-dollar warehouse operator. And in China, Guangzhou's Jianggao mail processing center has eight humanoids sorting parcels 24/7, currently at 800 parcels per hour, targeting 1,600 by second quarter 2027.
The common thread is that nobody's waiting for full autonomy. The winning play is deploy now, supervise, collect proprietary data, and let the models catch up. Whoever accumulates the most real-world manipulation data in the next 18 months probably wins this category.
US crypto regulation had a genuinely consequential week, and almost none of it was about legislation.
Start with the CLARITY Act. The Senate punted the vote to September. Majority Leader Thune says it'll be first in queue after recess, but passage still needs 60 votes, and Democratic support is soft heading into midterms. Former Defense Secretary Mark Esper made a notable pitch calling CLARITY a national security bill, arguing offshore crypto activity creates strategic vulnerabilities. Whether that framing moves any Senate votes is another question. Some in the industry are actually spinning the delay as positive, arguing a rushed bill with bad stablecoin provisions would be worse than no bill.
More interesting is what happened at the agency level. The SEC and CFTC signed a formal memorandum of understanding to coordinate on crypto, derivatives, and hybrid market rules. SEC Chair Paul Atkins publicly admitted that years of regulatory turf wars pushed activity offshore. The MOU covers product definitions, clearing and margin, cross-agency exams, surveillance, and enforcement coordination. It doesn't rewrite securities or commodities law, but it creates real machinery for handling disputes before they blow up into parallel enforcement actions. For firms building market infrastructure, exchanges, clearing, brokerage, this could genuinely lower the cost of operating in the US.
And then there's a live jurisdictional fight. The SEC granted CME's petition to review the conditional approval of Nasdaq PHLX's cash-settled Bitcoin index options, called QBTC. CME's argument is straightforward: Bitcoin is a commodity, these options are CFTC territory, and the SEC has no business approving them. The stay is in place, public comments are due by August 24. However this gets resolved will set precedent for every cash-settled crypto index product going forward.
Meanwhile, CFTC Chair Michael Selig says his agency has crypto rules ready to deploy the moment Congress gives statutory authority. So the pieces are lining up. Whether they actually snap together this year depends on the September Senate math.
One observation to end on. The bottleneck for AI in 2026 isn't compute anymore, it's memory. The bottleneck for humanoid robots isn't hardware, it's real-world operating data. And the bottleneck for US crypto markets isn't legislation, it's whether two agencies can actually cooperate. In all three cases, the constraint moved somewhere most people weren't looking a year ago. Watch the constraints, not the headlines.