Bitcoin is stuck in a holding pattern around 66,000 dollars, but the money keeps coming. Spot Bitcoin ETFs just extended their inflow streak to seven straight sessions, pulling in nearly 1 billion dollars, with BlackRock's IBIT doing most of the heavy lifting. In Washington, a new draft of the CLARITY Act dropped, and it would bar the President and other officials from issuing or holding crypto tokens through 2029. Perplexity made two big moves in AI infrastructure this week. BitMEX, the exchange that invented the perpetual swap, announced it's shutting down after 11 years. And a consortium led by Strategy and BlackRock pledged 15 million dollars to start quantum-proofing Bitcoin. Four stories worth your time today.
Perplexity dropped something interesting this week, and it's not another chatbot. They released a new orchestrator model for Perplexity Computer, their agent product, built on top of GLM 5.2 and post-trained for their specific harness. The pitch is simple: near-frontier performance at roughly one-third the cost of Opus 4.8. On Terminal-Bench 2.1, this thing scores 80, versus Opus at 76 and GPT-5.5 at 78. But the real story is the economics. Instead of running every agentic task through the most expensive frontier model, Perplexity built an escalation logic. Cheap base models handle the routine work. When something hard shows up, a built-in advisor escalates to a stronger model. On the WANDR benchmark, the advisor setup runs at about 2.1 times the GLM 5.2 baseline cost, while Opus-as-default runs at 6.1 times. That's roughly half the cost per task. This is a real philosophical shift. For the last two years, the default posture in agentic AI has been: throw the biggest model at everything and worry about the bill later. Perplexity is arguing that's dumb. Use the frontier model as an on-call consultant, not the full-time worker. And it lines up with what they announced alongside it, called SPACE. SPACE is the sandboxed runtime powering Perplexity Computer. Every task runs in its own disposable Firecracker microVM, destroyed at the end, with rolling snapshots so you can pause, resume, fork, or roll back agent sessions. They got sandbox creation time down from 185 milliseconds to 60, and 90th-percentile latency from 447 milliseconds to 89. Roughly 3 to 5 times faster than what they had before. Put the two announcements together and you see the actual product strategy. Cheap orchestration on top, fast disposable sandboxes underneath. If agents are going to run for hours, days, or months, unit economics matter more than benchmark bragging rights. Perplexity seems to be the first major player betting the company on that.
While everyone argues about chatbots, AI is quietly eating drug discovery. Three announcements this week that fit together. First, Latent Labs launched Latent-Y, an autonomous drug design agent. You type a plain-language goal, it handles target analysis, epitope selection, and refinement, and gives you lab-ready sequences. In published results across nine targets, it hit a 67% target-level success rate, with binding affinities in the single-digit nanomolar range. And they claim campaigns finish about 56 times faster than expert estimates. Every approved researcher gets 250 free designs a day. A UC Davis lab used it to design nanobody inhibitors for a difficult ion channel target from scratch, and most candidates worked on the first try. Second, there's Robin, an open-source agent from FutureHouse, Oxford, and Fordham. Robin proposes existing commercially available drugs for specific diseases by matching them to disease mechanisms. In a demo on dry age-related macular degeneration, it identified Ripasudil, a drug already approved in Japan for glaucoma, and lab experiments confirmed it nearly doubled the relevant cell activity. Drug repurposing done by an AI agent that reads the literature, designs experiments, and updates its hypotheses. Third, and this is the big infrastructure move, Microsoft committed 60 million dollars to the DOE's Genesis Mission. The idea is to link the 17 US National Labs into a single AI-driven scientific platform, with Microsoft's Azure and Foundry stack underneath. Carnegie Mellon, Argonne, and Lawrence Livermore are separately building an interconnected ecosystem of autonomous labs, using AI agents to generate robotic instructions and digital twins to simulate experiments before running them. The pattern here is worth noting. The frontier of AI right now isn't a better chatbot. It's an agent that designs a molecule, hands it to a robot, reads the result, and iterates. That loop, running in parallel across thousands of targets, is what actually shortens drug discovery timelines. It's not hype. UC Davis got working nanobodies on the first try.
Back to Bitcoin. Price is sitting around 66,000 dollars, up modestly on the week, and the market is clearly being propped up by ETF demand. Spot Bitcoin ETFs just logged their seventh straight session of net inflows, bringing the run to nearly 1 billion dollars. BlackRock's IBIT pulled in 505 million dollars over five days from July 14 to 20, averaging about 96 million per day. IBIT now holds around 737,000 BTC, worth roughly 49 billion dollars. On July 21 alone, IBIT took in 164 million, more than 80% of that day's total bitcoin ETF demand. The concentration is striking. And yet, price isn't moving much. Bitcoin is rangebound between 64,000 and 66,800 after a 13% recovery from July lows. Open interest is flat around 474,000 BTC, funding rates are barely positive, and the Fear and Greed index is at 34. The market is absorbing capital without exploding higher. The reason is a wall at 68,000 dollars. That's the short-term holder cost basis, according to Glassnode. It's five months of underwater buyers sitting at breakeven, waiting to sell into the first retest. Grayscale's head of research says Bitcoin may have already bottomed for this cycle and that macro factors, especially Fed decisions, now drive price action more than the halving-based cycle model. The Fed meets July 28 and 29. Markets expect a hold. But if the disconnect between spot demand and price doesn't resolve, this whale-led rebound starts looking fragile. Two things to watch. One, whether IBIT flows keep this pace, or whether we've seen peak institutional urgency for the month. Two, whether BTC can push through 68,000 without getting sold. If it does, the next leg is on. If it stalls, the range holds, and the story becomes: even record ETF inflows can't move price when there's a giant wall of breakeven sellers above. Also worth noting: BitMEX is done. The exchange that invented the 100x perpetual swap in 2014 announced it's shutting down September 23. The BMEX token crashed 90%. An era in crypto derivatives ends quietly, mostly because everyone else copied their product and did it better.
There's a quiet story unfolding in Texas that Bitcoin miners should be paying close attention to. Earlier this month, the US Department of Energy approved a 3.26 billion dollar loan to AEP Texas to modernize roughly 2,800 miles of transmission. The stated goal is to serve AI data centers, advanced manufacturing, and Permian Basin operations. AEP Texas is currently sitting on letters of agreement representing up to 41 gigawatts of potential new load through 2030. Now, LOAs aren't firm contracts. Many won't convert. But the direction is unmistakable. This is the third utility loan under the Energy Dominance Financing program. Federal capital is flowing directly into transmission upgrades that will be paid back by ratepayers over 30 years, and the beneficiaries are hyperscale AI loads that need firm, uninterruptible power. Here's the tension. For years, Bitcoin miners have thrived in Texas by being flexible. They curtail when the grid is stressed, they consume surplus renewable output, they act as demand-response resources. That flexibility was the pitch to ERCOT and to local governments. Abilene just approved a 200 megawatt Lancium project called Artemis that's marketed exactly that way, a controllable load resource that stabilizes the grid. But hyperscalers don't want to curtail. They want 24/7 uptime for training runs. And now the federal government is subsidizing the transmission buildout that lets them get it. That narrows the arbitrage window miners have relied on. Meanwhile, Bernstein analysts are arguing miners will have to strike deals with third-party AI providers just to stay relevant, essentially selling their power access and infrastructure to the AI buildout. Some are already doing it. Satoshi Energy signed two agreements to colocate Bitcoin and AI data centers at Texas wind farms, totaling 327 megawatts. The pure-play miner model is under pressure. And you can see it in the numbers. Olenox Industries reported June production down from May, with fleet utilization dropping from 81% to 70%, mostly because Texas summer heat forces them into low-power mode and curtailment. Miners eat that cost. AI data centers, on subsidized transmission, don't. The 41 gigawatt LOA figure at AEP Texas is the number to watch. If most of that converts, miners get squeezed out of the best interconnection queues. If it doesn't, flexible load stays valuable. Either way, the era where Bitcoin miners were the obvious buyer of stranded Texas power is ending.
One thing to notice today: Perplexity is trying to make agents cheap, and the DOE is making AI power expensive to compete with. Both bets are essentially the same wager, that AI workloads at scale are about to become the dominant economic force on the internet and on the grid. Bitcoin is either along for the ride or getting run over. Pick a side.