Bitcoin is hovering near $64,000 this weekend, clawing back some ground after Friday's sell-off, but the broader picture is ugly. Spot Bitcoin ETFs just posted a record $6.4 billion in outflows over 30 days. JPMorgan says roughly 20% of miners are now unprofitable, with all-in production costs running around $78,000 per coin. Meanwhile, AMD's new MI355X landed within 5% of NVIDIA on key training benchmarks, AWS is hinting it might start selling Trainium chips to third parties, and the SEC and CFTC just launched a joint review that could rewrite the rules for crypto derivatives. Lots to get through.
Let's start with the chip wars, because something genuinely interesting is happening. For years, the AI training conversation has been NVIDIA, NVIDIA, and more NVIDIA. That's still mostly true. In the latest MLPerf Training v6.0 round, NVIDIA GPUs showed up in 197 of 227 results. That's 87% coverage across all 7 workloads. AMD only submitted on 3 workloads and skipped four frontier ones entirely, including DeepSeek-V3 and Llama 3.1 405B.
But here's the part that matters. On the workloads AMD did submit, the new MI355X came within about 5% of an equivalent NVIDIA GPU. That is a real result. It's 3.5 times faster than last year's MI300X. AMD also rolled out a new FP4 training format called MXFP4 and a unified training framework called Primus. This is the closest AMD has been to credible parity in years.
And AMD is pressing the advantage on the CPU side too. They're pitching a rack-scale argument: forget per-chip benchmarks, look at usable compute per 100 kilowatt rack. By that math, AMD's EPYC 9965 delivers 2.37 times the throughput of NVIDIA's Vera CPU, and the upcoming Venice chip projects 3.30 times. The pitch is aimed squarely at agentic AI workloads, where orchestration, databases, and middleware are CPU-bound, not GPU-bound.
Then there's the Rackspace deal. AMD just signed a definitive agreement to deploy 30 megawatts of AMD compute across Rackspace data centers starting late 2026, expanding through 2028. The target is regulated enterprise workloads, healthcare in particular, where governance and accountability matter more than raw FLOPS.
And Amazon is sniffing around the same opportunity. AWS is reportedly considering selling its Trainium chips directly to third-party data centers, which CEO Andy Jassy hinted could be a $50 billion annual run rate if spun out. NVIDIA's moat is still enormous, but for the first time in a while, it's getting tested on multiple fronts at once.
Now to something that's been quietly accelerating. AI agents are showing up inside classrooms at scale, and not as experiments anymore.
FEU Tech in the Philippines just signed with OpenAI to become what they're calling an AI-native university. We're talking ChatGPT Edu and Codex access for over 14,000 students plus faculty, with custom AI agents for curriculum development, admissions, and enrollment. They're rolling out an AI Tutor based on faculty materials, an AI Classroom for interactive instruction, and AI-driven student success insights.
In Korea, the Busan Metropolitan Office of Education just launched BeAT, a generative AI tutor for students built on GPT-5.4 and OpenAI's image model. It does lesson planning, assessment generation, personalized learning plans, and multilingual translation. Teacher training runs through September.
Israel is going even further. With a teacher shortage biting hard, the country is rolling out AI-assisted English instruction to middle schoolers nationwide next school year. 180 schools, about 130 million shekels, roughly $45 million in budget. The framing is interesting: they're calling it personalized learning, not teacher replacement, but the trigger is clearly a staffing crisis.
Meanwhile the University of Sydney's Cogniti platform launched on Microsoft Azure Marketplace and is now deployed at Leiden University across four faculties, plus polytechnics in New Zealand. Nursing students run simulated clinical scenarios. Psychology students practice suicidal risk assessment. Pharmacy students rehearse customer interactions. The privacy story matters here: data stays inside the institution, not used for commercial AI training.
And Khan Academy rebuilt Khanmigo to live inside practice problems rather than in a sidebar chat, after a year-long pilot with 5,100 students. The design philosophy is interesting: hints before submission, step-by-step guidance after errors, with the explicit goal of preserving productive struggle while preventing students from spiraling on persistent mistakes.
The overreliance question is real. Independent studies on AI tutor effectiveness are mixed. But the deployment train has clearly left the station.
Bitcoin mining is in genuine pain right now. JPMorgan estimates the all-in cost to produce one BTC is around $78,000, while Bitcoin trades near $63,000 to $64,000. That's about 20% of miners running underwater. Hashprice sits around $33 per petahash per day. June saw a 10% drop in mining difficulty and a 12% fall in hashrate as operators toggled rigs offline.
In Q1 alone, public miners sold a record 32,000 BTC just to cover expenses. Total miner holdings are down to 1.8 million BTC from 1.86 million in 2023. They're bleeding the treasury.
May's production reports tell the same story with more texture. Bitdeer mined 921 BTC, up 370% year-over-year, but ended the month holding just 171 BTC. A year ago they held 1,351. They are aggressively cashing out mined Bitcoin to fund their AI buildout. Their AI Cloud business is running at about $69 million annualized with 90% GPU utilization across 4,248 GPUs, including new GB300 NVL72 clusters. Their flagship Tydal site in Norway is being converted to a 180 megawatt AI data center for NVIDIA Vera Rubin gear.
CleanSpark mined 671 BTC at 50 exahash, but their capacity has been flat since February. They're redirecting capital toward AI data centers in Georgia and Texas. They sold 404 BTC at spot and 250 via call options, ending up net positive only 17 BTC for May after option premiums. Average sale price was around $79,934.
BitFuFu mined 177 BTC, with self-mining finally surpassing cloud mining for the first time this year. Canaan self-mined 90 BTC but actual operating hash rate was well below guidance because a hosting agreement expired.
The pattern is unmistakable. The big public miners are pivoting hard to AI infrastructure, using their BTC stash as the funding source. That's a rational response to bad mining economics, but it raises a question: are these still Bitcoin companies, or are they becoming hybrid data center operators that happen to mine on the side?
And Texas, which has spent years courting miners with cheap power and sales tax breaks, is now flipping the script. Governor Abbott has told regulators to make data centers pay for the grid strain they cause. The free ride is ending.
Last topic, and it's a big one for market structure. The SEC and CFTC just did something they almost never do: they launched a joint, formal review of how crypto derivatives are regulated. Two simultaneous requests for public comment, 60-day window, focused on Title VII of Dodd-Frank.
The headline issue is perpetual futures. CME Group is currently suing the CFTC over the approval of crypto perpetuals, arguing those contracts are swaps, not futures, and therefore the CFTC approved them under the wrong framework. The joint review is essentially the regulators saying: fine, let's actually figure out where the lines are.
The scope is broad. They want input on definitions of swaps and security-based swaps, how to treat mixed swaps, how to classify event contracts covering elections, sports, economic indicators, and crypto markets. They're asking whether cash-settled perpetual contracts referencing equity securities should fall under one framework or another. They're asking whether firms could satisfy both agencies through coordinated compliance rather than dual filings.
SEC Chair Paul Atkins and CFTC Chair Michael Selig framed it as ending years of regulatory ambiguity. Whether that actually happens depends on what comes out the other end of the comment process, but the intent signal is clear. They want a unified framework rather than continued jurisdictional turf battles.
Why does this matter for Bitcoin holders? Because the classification decisions will determine which US venues can offer crypto perpetuals, what settlement and margin rules apply, and how competitive American exchanges can be against offshore platforms like Binance and Bybit. Right now, the most liquid crypto derivatives market on the planet operates almost entirely outside US jurisdiction. If this review produces actual clarity, that could change.
On the stablecoin side, three federal agencies, Treasury, the OCC, and the FDIC, have proposed rules that would essentially turn stablecoin issuers into pseudo-banks. Weekly confidential reports to the OCC, quarterly financials, full anti-money-laundering and sanctions programs. Good news for Tether and Circle who can absorb the compliance cost. Bad news for any smaller issuer hoping to compete.
Here's the thing worth noticing today. Public Bitcoin miners are selling coin to build AI data centers. That's not a hedge. That's a bet that compute for AI inference will be more profitable than mining Bitcoin for years to come. They might be right. They might be early. But if you're holding mining stocks expecting pure Bitcoin exposure, your thesis just changed under your feet.