Bitcoin just lost the 80,000 level, sliding to around 78,000 as a bond market revolt pushed Treasury yields above 4.55% and put a Fed hike back on the table. Spot Bitcoin ETFs bled 1 billion dollars in a single week, snapping a six-week inflow streak. Meanwhile, the CLARITY Act cleared the Senate Banking Committee, Kevin Warsh just took over as Fed Chair on his first day, and Strategy quietly admitted it might sell Bitcoin to repurchase 1.5 billion in convertible notes. Over in AI, Nvidia hit a record 219 dollars ahead of its May 20 earnings, Vera Rubin shipments are accelerating to July, and Anthropic shipped a feature that fundamentally changes how coding agents decide they're done.
Let's start with the price action, because this week told us exactly how tightly Bitcoin is still tethered to macro. Bitcoin spent most of the week trying to break above 82,000 on the back of the CLARITY Act passing the Senate Banking Committee in a 15-9 bipartisan vote. That bill would settle the SEC versus CFTC jurisdiction fight, classify most digital assets as commodities, and give DeFi developers and stablecoin issuers actual rules to follow. Santiment flagged a major spike of euphoria in sentiment. And yet, Bitcoin couldn't hold the level. By Friday it was under 79,000, and overnight into Saturday it slipped to 78,000, taking out roughly 500 million dollars in long liquidations. Solana and XRP dropped 5%. What happened? Two things. First, the 10-year Treasury yield ripped above 4.55%, and the market repriced Fed expectations hard. CME FedWatch now shows a higher probability of a 25 basis point hike than a cut, with March 2027 the most likely timing for the next move higher. April CPI came in at 3.8%, PPI at 6%. Roughly 62% odds now say no rate cuts at all in 2026. Second, ETFs. Spot Bitcoin ETFs shed 1 billion dollars in a week, including a single-day 635 million dollar outflow, the largest since January. Capital is rotating into AI stocks. And then there's the Warsh factor. Kevin Warsh started as Fed Chair this week, replacing Powell. He's viewed as more hawkish on inflation, though he's personally held crypto and has floated an AI-productivity thesis that would justify lower rates. Markets don't know which Warsh shows up at the June FOMC. One odd footnote: Arkham data shows over 1 billion dollars in Bitcoin has left wallets attributed to Bhutan in the past year. Bhutan says it doesn't recall selling any. Make of that what you will. If yields stabilize, the mid-80s come back into play. If they don't, the mid-70s are on the table.
Speaking of selling pressure, Strategy did something this week that deserves more attention than it got. On May 15, the company agreed to repurchase roughly 1.5 billion dollars in principal of its 2029 convertible notes for about 1.38 billion in cash. And in the Form 8-K, Strategy explicitly told investors it may fund the repurchase using cash reserves, ATM sale proceeds, or, and this is the part that matters, Bitcoin sale proceeds. Now, Saylor has been walking this line for a couple of weeks. Six days before Strategy's last buy, he floated the possibility of selling BTC, then pledged to buy 10 to 20 BTC for every coin sold and to end the year with more Bitcoin than it started. Then on May 11, Strategy bought 535 BTC for 43 million dollars at an average of 80,340. Almost the entire purchase, 533 of those 535 coins, was funded with common stock through the ATM program, not the preferred STRC vehicle that dominated April's buying. Total holdings now sit at 818,869 BTC, acquired for about 61.9 billion at an average cost of 75,540 per coin. That's over 3.9% of the entire 21 million BTC supply. So why is selling Bitcoin suddenly on the table? Math. Strategy has 8.2 billion in convertible debt and 1.5 billion in annual dividend obligations on its perpetual preferred stack. After adopting FASB fair value accounting, it posted a 12.5 billion dollar unrealized loss in Q1, which actually creates a tax-loss harvesting incentive. Selling some Bitcoin at a loss could shield gains elsewhere. The market read this as bearish, and you can see why. The largest corporate Bitcoin buyer publicly leaving the door open to sales, while ETFs are bleeding and macro is turning hostile, is not the setup bulls wanted. But the structural story hasn't changed. Strategy still has over 19 billion dollars of dry powder under STRC alone, plus capacity across STRD, STRF, STRK, and a potential 21 billion dollar MSTR share offering extension. The shift back to common stock funding this week tells you STRC demand has cooled since April. Saylor still wants 1 million BTC by year-end, which means roughly 24,000 BTC per month for the rest of 2026. The question isn't whether Strategy buys more. It's whether the financing machine keeps working when MSTR is down 59% from its 2025 peak and the equity premium compresses.
Now to AI, where the picture is the inverse of crypto: relentless demand, constrained supply, and money pouring in. Nvidia closed at a record 219.44 this week, briefly pushing the market cap above 5.3 trillion dollars heading into May 20 earnings. Fiscal 2026 revenue hit 215.9 billion, up 65% year-over-year. Q4 data center revenue was 62.3 billion, up 75%. Analysts have price targets near 270 and basically everyone expects beat-and-raise. The product roadmap is what's driving conviction. Nvidia confirmed Vera Rubin first shipments roll out in July to Microsoft, Google, Amazon, Meta, and Oracle, with mass production in the second half of 2026 on TSMC's 3-nanometer node. Each Vera Rubin rack is estimated at around 180 million dollars. Blackwell Ultra, the current generation, is targeting 50x performance and 35x cost reduction specifically for agentic AI inference. And the rental market tells you how tight things are. The Ornn Compute Price Index has Blackwell hourly rental at 4.08 dollars as of mid-April, up 48% in two months. Lambda, Crusoe, Nebius, FluidStack, all reporting sold-out Blackwell pods through Q3. The bottleneck is TSMC's CoWoS-L packaging capacity, which doesn't meaningfully expand until late Q3, plus a 20% jump in HBM3e memory pricing for 2026. Analysts now think the supply-demand gap persists until 2029. That's the worst compute shortage in five years. Which is exactly why AMD is suddenly being taken seriously as an alternative. The MI 450 and the new Helios platform are AMD's first real rack-scale, Nvidia-comparable offering, going head-to-head with Rubin. AMD also keeps gaining CPU server market share under Lisa Su, which matters because every AI rack needs CPUs too. It's not that AMD is going to dethrone Nvidia. It's that hyperscalers desperate for compute will buy anything that ships. When the rental floor is 4 dollars an hour and your pods are sold out through Q3, second-source procurement stops being optional.
And speaking of agentic AI, the coding agent space had a real moment this week. Two things happened that, taken together, change how we should think about autonomous software engineering. First, Artificial Analysis released the first cross-stack coding agent benchmark, evaluating not just models but the entire model-plus-harness setup. The winner: Cursor CLI running Claude Opus 4.7, scoring 61. OpenAI Codex with GPT-5.5 and Anthropic's own Claude Code came in close behind. Gemini CLI with Gemini 3.1 Pro came in last at 43. Here's the interesting part: when you hold the model constant, when both Cursor and Anthropic are running the same Opus 4.7, Cursor's harness outperformed Anthropic's own setup. The wrapper matters as much as the model. The cost story is even sharper. Cursor's Composer 2, a tuned model trained with reinforcement learning on agentic coding traces, runs about 7 cents per task. Other setups run up to 76 cents per task. That's a 10x cost gap for comparable performance. Composer 2 isn't a frontier model, it's a cost-efficient specialist. And on the leaderboards, Claude Mythos Preview just hit a perfect 100 on SWE-bench Pro, with GPT-5.3 Codex emerging as the best open-weight option at 77.3. Second, and more architecturally interesting, Anthropic shipped Claude Code 2.1.139 with a new slash-goal command. The core idea: separate the agent that does the work from the agent that decides the work is done. The developer defines a completion condition, things like all tests pass, or lint clean, and a separate evaluator, by default Claude Haiku, checks after each step whether the goal is met. If not, the worker keeps going. The framing from the team behind it is blunt: you can't trust a model to judge its own homework. This is the bigger insight. The original SWE-agent research from NeurIPS 2024 showed that giving GPT-4 a proper terminal interface, what they called an Agent-Computer Interface, moved patch success on real bugs from roughly 1% to double-digits. The lesson was that interaction design beats prompt engineering. Now we're seeing the next layer: separating execution from evaluation. Coding agents are starting to look less like clever prompts and more like actual distributed systems with workers, judges, and audit logs.
Here's the prediction. The same week Bitcoin's largest corporate holder admitted it might sell coins to service convertible debt, hyperscalers signed up for 180 million dollar AI racks they can't get delivered until July. Capital is voting, loudly, on where the next decade's productivity comes from. Bitcoin still has to earn its bid back the hard way: through legislation, through lower yields, and through buyers who don't need to sell.