Bitcoin's clawing its way back above $77,000 after holding the $75,000 line, capping its best month in a year. But under the hood, this rally was futures-driven, not spot-driven, and that's making analysts nervous. Meanwhile, the AI funding spigot is wide open. DeepMind veteran David Silver just raised $1.1 billion for a lab that wants to build AI without human data. Legal AI startup Legora hit a $5.6 billion valuation. And GitHub is killing flat-rate Copilot pricing — usage-based billing kicks in June 1. Riot Platforms posted a messy quarter but pivoted hard into AI data centers with AMD as its anchor tenant. And Metaplanet is sitting on roughly $490 million in unrealized Bitcoin losses while still trying to buy more. Let's get into it.
Bitcoin closed April above $76,000, its best monthly gain in a year, and it's now trading just over $77,000. On the surface, that's a win. The S&P 500 also hit fresh all-time highs, big tech earnings came in strong, and spot Bitcoin ETFs pulled in $2 billion in April — the biggest monthly inflow of the year so far, with BlackRock's IBIT leading.
But here's where it gets uncomfortable. CryptoQuant flagged that April's rally was almost entirely futures-driven. Spot demand actually shrank. That's the same market structure we saw in 2022 — leverage-fueled rebounds that gave way to fresh downside because there were no real buyers underneath. Open interest is flat, funding is negative, and traders keep a short bias even as price grinds up. Every push toward $77,000 gets sold. Profit-taking is capping rallies before they can test $80,000.
Late in the week, spot ETF flows actually flipped — over $490 million in outflows hit the tape as oil prices climbed and AI growth metrics disappointed. So the institutional bid that drove April is showing cracks right at the top.
The bullish counter-narrative came from Ark Invest, which doubled down on a $16 trillion Bitcoin market cap target by 2030, driven by institutional demand. That implies a price north of $700,000 per coin. It's a fine long-term thesis, but it doesn't change what's happening in the next two weeks. The $75,000 cost basis cluster is the line in the sand. Hold it, and the structure stays intact. Lose it, and the 2022 analog gets a lot more interesting — and not in a good way.
One quieter data point worth flagging: Bitso reports stablecoins have now overtaken Bitcoin in Latin American crypto purchases. In inflation-hit economies, people want dollars, not volatility. That's a structural shift in how crypto is actually being used at the retail level, even as the institutional Bitcoin narrative dominates the headlines.
The AI funding numbers this week are absurd, and they're telling us something about where smart money thinks the next decade goes.
Start with the biggest one. David Silver — the DeepMind researcher behind AlphaZero, the program that taught itself chess and Go from scratch — just raised $1.1 billion at a $5.1 billion valuation for a new lab called Ineffable Intelligence. Sequoia and Lightspeed led, with Google and Nvidia participating. The pitch: build a superlearner that discovers knowledge through reinforcement learning, with zero human data. If you believe large language models are hitting a ceiling because they've already eaten the internet, this is the bet on what comes next. Silver's also pledging personal proceeds to charity, which is a nice touch but not why Sequoia wrote the check.
Legal AI is its own gold rush. Legora, the Swedish startup, raised a $50 million extension to its Series D, hitting a $5.6 billion valuation. They crossed $100 million in ARR and serve over 1,000 law firms. Their main competitor, Harvey, claims roughly 100,000 lawyers across 1,300 organizations. And then there's Manifest OS, which raised $60 million at a $750 million valuation — a different model where they actually recruit lawyers into a network and give them AI tools to bill more. Three big bets, three different theories of how AI eats the legal profession.
Parallel Web Systems — Parag Agrawal's post-Twitter venture — raised $100 million at a $2 billion valuation, just five months after their last round at $740 million. They sell web search APIs built specifically for AI agents. Customers include Harvey, Notion, and Clay. The signal here is that the agent infrastructure layer is consolidating fast, and Sequoia keeps writing bigger checks at higher prices.
And a smaller one worth mentioning: Standard Intelligence, six people, raised $75 million from Sequoia and Spark with Andrej Karpathy angeling. They're building computer-use models trained on 11 million hours of video, claiming a video encoder 100 times more efficient than OpenAI's. If that holds up technically, it's a genuinely interesting efficiency play in a field where everyone else is just throwing more GPUs at the problem.
GitHub announced this week that Copilot is moving to usage-based billing on June 1. This is a bigger deal than the headlines suggest, and it tells you exactly how the economics of agentic AI are shaking out.
Here's the structure. Base subscription prices stay the same — Pro is $10, Pro Plus is $39, Business is $19 per user, Enterprise is $39 per user. But every plan now comes with a monthly allotment of GitHub AI Credits, and credits get burned by token consumption at the underlying model's API rates. Run out, and you either buy more or you stop. Code completions and next-edit suggestions stay free and don't touch your credit balance. But anything agentic — multi-step tasks, cloud agents, code review — that's metered.
Why now? Because the old per-seat model was getting destroyed by agentic workloads. When a single Copilot session might run an autonomous agent for an hour, planning, executing, and iterating across dozens of files, the token cost per user goes from cents to dollars. GitHub was eating that delta. Now they're not.
The pricing also lands right as GPT-5.5 hits Copilot for paid tiers. It's a purpose-built agentic coding model — 82.7 percent on Terminal-Bench 2.0, 58.6 percent on SWE-Bench Pro. But it carries a 7.5 times premium multiplier. So your shiny new agent is also the one that drains credits fastest.
Meanwhile, Cognition — the company behind Devin, the autonomous AI software engineer — is reportedly in talks to raise hundreds of millions at a $25 billion valuation. That's up from $10.2 billion in September. Devin competes directly with Copilot's agent mode, and the funding is the market voting on whether autonomous engineering agents are real or hype.
The pattern across all of this: per-seat SaaS pricing is breaking under agentic load. Every vendor selling AI tools to developers is going to face the same metering reckoning GitHub just had. Buyers should plan for it. Your AI tooling line item is about to look a lot more like an AWS bill than a Slack subscription.
Two stories that are really one story: what happens to Bitcoin-leveraged businesses when the price stops going up.
Riot Platforms reported Q1. Revenue beat at $167 million, but the GAAP loss was $500 million — driven by $327 million in non-cash Bitcoin mark-to-market losses. They mined 1,473 BTC at a direct cost of $44,629 per coin, down 26 percent year-over-year. They hold 15,679 BTC worth about $1.1 billion. Fine numbers, ugly headline.
The real story at Riot is the AI pivot. AMD exercised its option for an additional 25 megawatts at the Rockdale facility, bringing total contracted capacity to 50 megawatts under a 10-year, $636 million deal. Projected average annual NOI from that footprint is around $51 million. Data center revenue hit $33 million this quarter — most of it from one-time tenant fit-out fees, but recurring lease revenue is supposed to hit $37.8 million by year-end and $55.6 million once AMD's full 50 megawatts come online in 2027. CEO Jason Les called it an inflection point. He's not wrong. Riot is becoming a hyperscale landlord that happens to mine Bitcoin on the side. Expect every public miner to attempt some version of this pivot.
Then there's Metaplanet. The Japanese Bitcoin treasury holds 40,177 BTC at an average cost somewhere between $94,000 and $104,000 per coin. With Bitcoin near $77,000, that's roughly $490 million in unrealized losses. The stock is down 22 percent year-to-date and trades at a 36 percent discount to its Bitcoin holdings — about $2 billion market cap against $3.1 billion in BTC.
And they keep buying. This week, Metaplanet raised another $50 million via zero-coupon bonds, fully underwritten by EVO Fund. Their target: 100,000 BTC by end of 2026 and 210,000 BTC — about 1 percent of all Bitcoin — by end of 2027. To stay on pace, they need to add roughly 60,000 more coins this year alone.
The risk is the reflexivity. Metaplanet's revenue model relies on selling options against its Bitcoin stack and issuing shares to buy more. If the share price stays below NAV, they can't issue equity efficiently, which throttles the whole flywheel. They're now reportedly buying ads on the Las Vegas Sphere to lift sentiment. When your treasury strategy depends on advertising spend, something has shifted. Strategy figured out how to make the discount-to-NAV problem work in their favor through years of execution. Metaplanet is finding out the playbook doesn't auto-port to a different market in a different cycle.
One question to chew on. If GitHub couldn't make per-seat pricing work for Copilot, and per-seat is the most successful business model in software history, what does that say about every other AI product still pretending the old economics apply?