Bitcoin is sitting on a knife's edge below $60,000, with $4 billion in ETF outflows last month and Strategy openly talking about selling Bitcoin to fund dividends. AMD's stock is up 144% this year while Nvidia is barely up 4%, even though Nvidia still owns the data center. El Salvador's Bitcoin pile keeps growing, but the IMF says they haven't actually bought any. And a New York school district is putting a humanoid robot tutor in front of 500 high schoolers this fall. Let's get into it.
Bitcoin spent the last week pinned in a tight band between $59,000 and $60,000, and that quiet is starting to look ominous. The token broke below $60,000 in Asian trading as the Japanese yen hit a 40-year low against the dollar, lifting the dollar across the board and squeezing risk assets. Options traders are paying up for downside protection. The put-call ratio hit a one-year high. CME open interest is dominated by puts betting against the $60,000 floor, and the next clean target if support breaks is somewhere between $54,000 and $56,000, with $40,000 floating around as the worst-case scenario.
The institutional bid that carried this cycle has gone quiet. June saw $4.06 billion in net outflows from U.S. spot Bitcoin ETFs, the worst month since they launched in January 2024. BlackRock's IBIT alone shed about $300 million in a single session and roughly $1.3 billion across the week. Strategy is heading for its 11th losing month out of the last 12, with MSTR down about 41% in June. And then Strategy dropped its new capital framework: a $2.55 billion dollar-denominated reserve to backstop the STRC preferred dividend, $2 billion in buybacks authorized, and crucially, board approval to monetize Bitcoin if needed to fund those obligations. Benchmark called it bullish and slapped a $570 price target on MSTR. The market's reaction was more mixed, because for the first time, Saylor is openly telling investors he might sell coins.
There's a counter-narrative worth holding onto. Swan's Cory Klippsten points out that long-term holder supply just hit a record 16.65 million BTC, while short-term holders are bleeding coins. That's classic late-stage capitulation behavior, the kind that historically marks bottoms, not tops. MidChains says sovereign funds are treating this discount as an entry point. Silicon Valley Bank is calling this the start of a real institutional Bitcoin lending era, with stronger risk controls than the 2022 blowup. Friday's $10.6 billion quarterly options expiry is the next real test. Reclaim $68,000 to $70,000 and the gamma flips back bullish. Lose $60,000 cleanly and the hedges start paying off.
AMD stock is up 144% in 2026. Nvidia is up about 4%. If you stop there, it sounds like AMD finally caught Nvidia in AI. The numbers say otherwise. Nvidia's data center revenue last quarter was around $75 billion, growing 92% year over year. AMD's data center revenue was about $5.8 billion, growing 57%. Nvidia is still roughly 13 times the size in the segment that actually matters, and growing faster off a bigger base.
The story everyone's pricing into AMD is Helios, the new rack-scale platform built around the Instinct MI455X. AMD will unveil it at the Advancing AI keynote on July 22nd and 23rd. The pitch is real: 432 GB of HBM4 per GPU versus 288 GB on Nvidia's Vera Rubin, roughly 50% more memory per rack. That matters for training huge models and for inference workloads that hate cross-rack parallelism. Oracle Cloud is already lined up for a 50,000-GPU Helios rollout in Q3. But here's the catch: Helios is only going into limited production in the second half of 2026. Mass production is a 2027 story. Nvidia's Vera Rubin NVL72 is shipping right now. And AMD had to fall back to UALink-over-Ethernet for the interconnect because the full UALink ecosystem isn't ready at scale.
The more interesting argument is about Nvidia's moat. CUDA used to be unassailable. It's still dominant, but the cracks are visible. PyTorch with Triton lets you skip CUDA on inference. OpenAI's internal stack is actively trying to drive inference costs down on alternative hardware. Anthropic has built a serious business on Google's TPUs and Amazon's Trainium without touching Nvidia for a lot of its workload. Rubin Ultra's design got cut from 4 chips to 2 because of CoWoS packaging constraints. And here's the valuation gap nobody's solving: AMD trades at 71 times forward earnings. Nvidia trades at about 21 times. If Nvidia sentiment normalizes even slightly, it's the one that re-rates higher, not AMD. The market is paying a Nvidia-killer multiple for a company that's still a distant second.
El Salvador's Bitcoin reserve sits at 7,696 BTC, worth about $461 million. That's up from roughly 5,968 BTC at the end of 2024. Bukele's government keeps publishing the running total. The catch: the IMF says no new market purchases have actually happened. The increases are coming from consolidating Bitcoin between government wallets, including a cold storage address held by the development bank BANDESAL. Under the Extended Fund Facility loan, El Salvador agreed to a zero ceiling on voluntary public sector Bitcoin purchases. So the famous "one Bitcoin a day" rhetoric and the actual on-chain accounting don't quite match. Bitcoin lost legal tender status back in January 2025. The Chivo wallet is being wound down. The maximalist experiment as originally pitched is essentially over.
And yet, the reserve keeps growing on paper, and that's the interesting part. Whether it's market buys or internal shuffling, the political signal is the same: a sovereign treasury can hold Bitcoin without forcing it on merchants, without breaking IMF conditionality, and without doing anything market-moving. It's the quiet model. Compare that to Bhutan, which used surplus hydropower to mine its way to over 13,000 BTC at peak, fourth largest government holder behind the US, China, and the UK. Then in 2023 they sold $100 million worth to double civil servant salaries. From late 2024 on, they've sold down more than 70% of holdings. The Bhutan story shows the other side: if you actually need the money, sovereign Bitcoin gets liquidated like any other reserve asset.
Meanwhile in Washington, Senator Cynthia Lummis is still pushing S.954, the bill that would direct the Treasury to buy up to 1 million Bitcoin over five years and hold it for at least 20 years, modeled on the gold reserve. That's 5% of total supply held by one government. Lummis argues it could meaningfully dent a $36 to $39 trillion national debt over two decades. The concentration risk is obvious. So is the political risk that a future administration just sells it. But the direction of travel is unmistakable: Bitcoin is quietly being normalized as a sovereign reserve asset, even by governments that publicly walked back their experiments.
Salamanca City Central School District in upstate New York is about to put a humanoid robot in front of about 500 high schoolers this fall. Realbotix is deploying its M-Series robot paired with the Optio AI tutoring system. Optio handles district-approved curriculum tutoring, multilingual homework help, and 24/7 one-on-one support. The robot itself adds face-to-face interaction, facial expressions, and real-time conversation, the embodied AI piece that's supposed to keep kids engaged in ways a chatbot on a Chromebook doesn't. Students in the Woz ED AI and Robotics courses get hit first, then it scales to the broader high school population. It's one of the first deployments of its kind in a U.S. public school, not a pilot, not a demo, an actual classroom tool.
At the same time, Google and Microsoft are racing to put more capable agents directly into the software students already use. Google rolled out Gemini Study Notebooks: students drop in materials, take a diagnostic quiz, and get a personalized study plan that updates as they progress, with free full-length GRE and ACT practice tests through The Princeton Review. SAT support is being added. It syncs with NotebookLM and is heading into Google Classroom for school accounts. Microsoft answered with new Microsoft 365 Education features including Teach in Copilot, which generates standards-aligned unit plans for over 50 countries, and a Study and Learn Agent built around learning science, with scaffolded questions and immediate feedback for students 13 and up, all included with Microsoft 365 Education at no extra cost.
The interesting tension is between the two models. The Realbotix deployment is high-touch, embodied, deeply integrated into a single district. The Google and Microsoft approach is mass-market software at zero marginal cost, and 87% of educators say they want training on how to use it. Both are betting that the bottleneck in education isn't content delivery anymore. It's personalization at scale. The honest open question is what happens to the students who use these every day for four years. We don't have that data yet. We're about to get it.
If Bitcoin loses $60,000 cleanly this week, watch what the sovereigns do, not what the ETF tape does. The retail and institutional bid that drove this cycle has clearly cooled. The next bid, if it shows up, will look quieter, slower, and a lot more patient.