Microsoft just made Copilot Cowork generally available, turning Office 365 into a platform for autonomous, multi-step AI work with usage-based billing. Foxconn and NVIDIA are deploying agentic AI across Taiwan's biggest hospitals, and a Nature paper introduces MIRA, an AI agent that matched physicians in full emergency department workflows. Bitcoin is sitting near $64,000 after the Fed held rates but turned hawkish under new chair Kevin Warsh. And Foundation, the Passport hardware wallet maker, raised $6.4 million led by Fulgur Ventures to push into AI security and agent authorization. Let's get into it.
Microsoft Copilot Cowork is now generally available worldwide, and this is a meaningful shift in how enterprise AI gets sold. Cowork is a cloud-based agent that runs inside your Microsoft 365 tenant, executes long-running, multi-tool tasks even when your laptop is off, and returns completed work, not drafts. It can pull from your documents, your meetings, Teams, email, and the web through Edge. Sensitive actions like sending mail or scheduling still require human approval, and it inherits the signed-in user's permissions.
The interesting part is the pricing. Cowork sits on top of the standard Copilot license, which runs about $30 per user per month at the enterprise tier, and adds usage-based billing through Copilot Credits. You pay 1 cent per credit on pay-as-you-go, or commit to volume for a discount. Four things drive cost: which model you use, how much context it pulls, how many tool calls it makes, and how long it runs. Available models include Anthropic's Opus 4.8 and Sonnet 4.6, and for Frontier customers, GPT-5.5 and Microsoft's own Cowork 1. Microsoft is also exploring a DeepSeek option to push costs down.
Admins get real controls — Cowork is off by default, with tenant, group, and user spending limits, alerts, and per-task cost visibility coming soon. Microsoft is claiming 30 to 40 percent cost savings per prompt versus running Claude's own Cowork product inside their environment. Take that with salt — it's an internal benchmark.
The bigger story: Microsoft just normalized metered AI billing for knowledge work. Seat licenses don't capture agent value, and they're admitting it. Every enterprise software vendor is watching this pricing model, because if it sticks, the per-seat era starts ending right here.
Healthcare AI crossed a real threshold this week, on two fronts.
First, Foxconn and NVIDIA announced major agentic deployments across Taiwan's largest hospitals — Chang Gung Memorial, NTU Hospital, Taipei Veterans, and others, covering more than 14 million patient encounters a year. This is part of Taiwan's $1.5 billion Healthy Taiwan program. Foxconn's CoDoctor platform aggregates AI agents for cardiology, oncology, and ophthalmology. CoDoClaw extends that with NVIDIA's NemoClaw for multi-agent orchestration. There are physical agents too — Nurabot nursing robots handling logistics, and a Scrub Bot assisting in live surgeries. Foxconn says using Omniverse digital twins to simulate hospitals first cut deployment time by 40 percent with 98 percent navigation accuracy.
Second, and more striking, Nature published a paper on MIRA — Medical Intelligence for Reasoning and Action. MIRA is an autonomous agent running inside a sandboxed electronic health records environment with 11 tools. It chats with a patient agent, orders labs and imaging, generates differential diagnoses, prescribes, admits, and schedules. Researchers tested it on more than 500 MIMIC-IV emergency department cases against board-certified physicians. MIRA matched or beat doctors on diagnostic accuracy and treatment quality, with strong medication safety on renal dosing, interactions, allergies, and opioid risk.
And Cleveland Clinic Abu Dhabi launched Aila, built with Owkin — pitched as the world's first clinical AI scientist, starting with prostate cancer, pulling from EHRs, physician notes, pathology, and imaging.
The pattern is clear: medical AI moved past chatbots and copilots in 2026. These are agents taking structured actions inside clinical systems. The MIRA result especially — physician-level performance on the full ED workflow — is the kind of evidence regulators and hospital systems were waiting for.
Bitcoin is around $64,000, having printed a week-to-date low of $64,500 and briefly dipping under the 200-week moving average twice in the past two weeks. The Fed held its target range at 3.50% to 3.75% on June 17, Kevin Warsh's first meeting as chair. But the updated dot plot turned hawkish — 9 of 18 members now project at least one hike before year-end, against 8 holding. The statement signaled the committee is more worried about inflation than growth. The dollar index is testing a major breakout. Spot Bitcoin and Ether ETFs saw $111 million in combined outflows.
The pressure is showing up in the corporate-treasury complex too. Strategy's STRC preferred stock — the variable-rate perpetual Saylor uses to fund Bitcoin purchases — hit a record low below par, around $91.79. Bearish options positioning is building, and STRC dividends already forced Strategy's first BTC sale this month. That's a structural concern: when the funding vehicle weakens, the buying engine slows.
Not all bearish though. Glassnode says recent capitulation was twice as weak as prior cycles, with realized losses down 46% as spot liquidity turns supportive. Kraken notes that historically, buying Bitcoin below its 200-week average has delivered median returns over 100%. Bitwise's Matt Hougan argues the next bull run will be slower and less volatile as investor appetite shifts toward stablecoins and tokenization. And Capital B's shareholders just approved up to $120 billion in financing capacity for Bitcoin accumulation — so the treasury thesis is far from dead, just stratifying between the well-capitalized and the over-leveraged.
The takeaway: macro is the dominant variable right now. A hawkish Fed plus a breaking-out dollar is a tough setup for Bitcoin, but the on-chain damage looks contained.
A few notable raises this week, and one that pulls these themes together nicely.
Foundation, maker of the Passport hardware wallet, raised $6.4 million led by Fulgur Ventures, with Arche Capital participating, bringing total funding to about $16.5 million. The interesting move: Foundation is explicitly expanding beyond Bitcoin custody into AI security infrastructure — agent authorization, identity verification, hardware-based multi-factor approval. Their thesis is that as AI agents start moving money and signing transactions on behalf of users, you need an independent hardware layer to approve those actions. Passport Prime is now broadly available, and they're opening a KeyOS app marketplace later this year, with Cake Wallet as the first external developer. This connects directly to the MoonPay-Ledger integration that went live June 16, where MoonPay Trade now runs cross-chain swaps across 200-plus blockchains directly from the Ledger device, with every trade signed on hardware. Same pattern — hardware as the trust anchor for an increasingly automated stack.
Meanwhile Trezor's CCO Danny Sanders used the BTC Prague stage to warn that ETFs are quietly eroding self-custody. Of roughly 600 million crypto users globally, only about 10% self-custody, and just 12 to 13 million use hardware wallets. Over $53 billion has flowed into US spot Bitcoin ETFs since 2024. Sanders called the assumption that you can just buy an ETF the worst-case scenario for the industry.
Other raises worth noting: stablecoin compliance startup Range pulled $8.3 million Series A from TX Ventures, Maven 11, and others. Trace Finance raised $32 million for cross-border stablecoin settlement. And Paradigm led a $9 million Series A in El Dorado, a Latin American stablecoin payments platform operating in 12 countries on the Tempo blockchain.
The through-line: capital is flowing to infrastructure that sits between users, AI agents, and traditional finance rails. Wallets becoming agent-approval devices. Compliance layers becoming pre-execution gates. Stablecoin rails replacing correspondent banking. Less speculation, more plumbing.
Watch how fast hardware-approval and metered-agent billing converge. The companies that figure out how to charge for autonomous work and prove it's authorized by a human will own the next platform layer.