Bitcoin punched back above 65,000 on twin inflation misses before fading to 64,000 as Iran struck US bases again. Anthropic dropped Claude Science, a full flagship product for computational biology and drug discovery, and made Claude free for US teachers through mid-2027. Marathon signed a 600 million dollar deal for a Texas mega-site targeting 2 GW by 2028. And the US Senate unanimously told the White House: no clemency for Sam Bankman-Fried. Four stories worth your time today.
Anthropic just made its most interesting product move in a while. Claude Science launched this week as a standalone flagship, available to all paid subscribers, and it's aimed squarely at scientific research the way Claude Code is aimed at software engineering.
The pitch is specific. Claude Science can write code, run it on compute clusters, interface with genetics, chemistry, and protein biology tools, and hold itself to reproducibility standards that actual scientists will accept. In the launch demo, it autonomously identified drug candidates for phenylketonuria, a rare metabolic disease. Anthropic says it plans to use the tool internally for its own drug research into neglected diseases, and it's openly courting pharma partnerships.
Read between the lines and this is Anthropic aiming directly at DeepMind's turf. AlphaFold gave Google's AI arm a moat in biology, and Anthropic has been quietly hiring from that world. There's also a very practical angle here: Anthropic is reportedly moving toward an IPO, and pharma contracts are a much cleaner revenue story than token-metered chatbot subscriptions.
On the same day, Anthropic launched Claude for Teachers, giving verified US K-12 teachers free access to the full premium stack, including Claude Code and Cowork, through June 2027. The interesting technical piece is Learning Commons, a retrieval-augmented system grounded in a live knowledge graph of roughly 150,000 state curriculum standards. Lessons get scaffolded against the actual standards for your state, not hallucinated from training data, and the standards update without retraining the model. Teacher and student data doesn't get used for training.
Two very different products, same underlying bet: general-purpose chatbots are getting commoditized, and the money is in domain-specific agents that plug into real workflows with real ground truth. Claude Code proved it for developers. Now Anthropic is running the same playbook on biology labs and classrooms.
Bitcoin's had an interesting 48 hours. CPI came in soft on Monday. Then Tuesday's PPI dropped 0.3 percent month-over-month, the sharpest wholesale price drop since August 2025, when the market was looking for flat. Two consecutive inflation misses, and the short book got torched.
About 100 million dollars of shorts got liquidated in the 30 minutes after the PPI print. The day before, on the CPI release, 134 million in shorts went in an hour. Total liquidations for the run topped 324 million, with shorts making up around 85 percent of that. Bitcoin printed a monthly high of 65,518 before pulling back to the low 64s after fresh Iranian strikes on US bases.
CME FedWatch now puts the odds of a July rate hike at roughly 10 to 13 percent. The July 28-29 FOMC meeting is the next real catalyst.
But here's where it gets more interesting. Glassnode is calling the buyers at 107,000 back in the earlier leg of this cycle the early signals of a 2026 bear-market bottom, with 69,000 emerging as the new price battleground. And Galaxy Research is flagging that long-dormant coins, coins aged one year or more, have gone quiet in 2026. On-chain movement of old coins is less than half of 2025's pace. Alex Thorn reads that as old hands stopped selling and new hands now setting the marginal price.
Also worth noting: a wallet dormant since the 2017 peak just moved 383 million dollars in bitcoin, but to a fresh address, not to an exchange. Nothing sold, yet.
And Bitwise flagged the divergence of the year: publicly traded crypto companies gained 23 percent in the first half of 2026, while crypto tokens fell 36 percent. That's a 59-point gap. Either the equities are pricing in a recovery the tokens haven't priced yet, or the equities are capturing the fee-and-services revenue that adoption generates regardless of spot price. Both stories can be right.
Two datapoints this week that put some cold water on the autonomous AI trading narrative.
HKU Business School just released results from Agentic Trader, a six-week live FX experiment running ten frontier models with identical capital, tools, leverage, and real-time market data. No predefined strategies. Fully autonomous. The results are messy in the way real markets are messy.
Alibaba's Qwen won, roughly 10 percent return. Kimi and Seed also posted strong numbers. GPT and GLM models hovered near break-even. Claude and MiniMax underperformed. DeepSeek came in dead last at negative 15.1 percent, with the biggest drawdown of the group. Trade counts ranged from 200 to over 1,000, and higher activity did not correlate with better returns. Leverage discipline mattered more than model IQ.
The key finding: static benchmarks don't predict live-market performance. Adaptability and risk management do.
Parallel to that, a new paper introduced TradeLens, a framework asking a sharper question: can agentic trading systems pay for their own intelligence? Meaning, does the LLM inference cost actually get earned back by decision quality? Because if you're burning tokens on thousands of trades that don't add alpha, you don't have a trading system, you have an expensive random number generator. Different backbones showed different failure modes: DeepSeek struggled with asset selection, GLM struggled with timing.
Meanwhile the capital keeps flowing. Grace Investment Machine closed a 20 million dollar Series A, its third round in its first year, for autonomous agentic investing. Aethon Fund launched with 50 million and a transparency pledge. Alpaca raised 135 million to build tokenized, agent-first brokerage infrastructure. And Visa and Artemis published research arguing the autonomous AI agent economy is still bottlenecked by payment rails, which is why the x402 Foundation just convened as a neutral standards body for AI agentic commerce.
The money is betting on the thesis. The live data says most models still can't beat a coin flip in FX. Both can be true for a while.
The public Bitcoin miners are finishing their transformation into power-and-compute companies, and this week made that unmistakable.
Marathon Digital filed on a deal for a Matagorda County, Texas site: over 1,200 acres, up to 1 GW of grid capacity by October 2027, and up to 2 GW by April 2028 pending ERCOT approval. Payments are milestone-based, capped at 600 million dollars. Combined with the pending 1.5 billion dollar Long Ridge acquisition in Ohio, MARA is putting roughly 2.1 billion in capex on the table, close to 44 percent of its market cap. The stock actually dipped after the filing, down about 5.5 percent, because there's no signed tenant lease yet. The market is pricing execution risk, not vision.
Riot Platforms told a similar story with Q1 numbers. Total revenue 167 million, but bitcoin mining revenue fell to 112 million as network hashrate climbed 24 percent and BTC softened. The data center segment brought in 33 million, engineering services another 22 million. Riot sold over 250 million in bitcoin for treasury management, produced 1,473 BTC, and still holds 15,679 BTC on balance sheet. It doubled its contracted AMD data center capacity from 25 MW to 50 MW.
CleanSpark went even further out on the plank, signing a 6.6 billion dollar AI lease before securing the 2.1 billion in financing needed to build the 175 MW site. Long revenue runway, undisclosed funding terms, and rent that can be reduced or terminated if milestones slip. Bold.
And then the counterweight: on July 14, New York Governor Kathy Hochul signed an executive order pausing all incomplete permit applications for new large data centers, including a specific 50 MW project. First statewide moratorium in the country, and it lands right as miners are trying to reposition as AI hosts. Power is now the constraint, and regulators are starting to notice.
The pattern is clear. The old miner story of "stack sats, sell when needed" is dead as a standalone thesis for public companies. The new story is: control gigawatts, sign hyperscaler leases, and let bitcoin mining be the flex load. Whether the equity market rewards that or punishes it depends entirely on tenant signatures showing up in the next few quarters.
One prediction to close on. The next twelve months, the winning bet in this space isn't which token pumps or which model tops a benchmark. It's who signs the contracts. Anthropic signing pharma. Marathon and CleanSpark signing hyperscalers. Cantor and Securitize signing issuers for onchain IPOs. Vision is cheap right now. Signatures are what get repriced.