Monday, July 27, 2026. Bitcoin is back above $65,000 after the U.S. and Iran paused strikes near the Strait of Hormuz, sending oil down 5% and lifting risk assets across the board. Ether is outperforming, up 4% to nearly $1,964. Traders are stripping off downside hedges heading into Wednesday's Fed decision, with the put-call ratio collapsing from 0.76 in late June to about 0.52. Meanwhile, BitMart shocked the market by abruptly winding down after nine years, freezing withdrawals and echoing the FTX-era chaos of 2022. Circle just bought nearly 1,000 blockchain patents from IBM. And a coalition led by NVIDIA, Microsoft, and Meta is pushing Washington to protect open-weight AI, with OpenAI joining the letter days later. Anthropic did not sign. Four stories today: Meta's frontier push with Llama 5 and Muse Spark, OpenAI Presence taking over call centers, Lightning Labs launching Wavelength for agents, and the growing gap between Tether and Circle under the GENIUS Act.
Meta is having a moment. Llama 5 dropped at the end of June, and the numbers are legitimate: 88.5 on MMLU-Pro, 79.6% on SWE-bench Verified, a 1M-token context window, and coverage across 40-plus languages. Over 1.1 million developers pulled it down in the first 72 hours. Total Llama downloads are past 400 million. This is the widest open-weight generalist model on the market, and enterprises are eating it up because they can run it on-prem, air-gapped, fine-tuned on proprietary data, with no vendor lock-in and no per-token bill to a cloud provider. Meta has spent around $40 billion on AI infrastructure through mid-2026, including their MTIA 3 chips and two new hyperscale data centers in Louisiana and Iowa pushing over 2 GW of capacity. But here's the interesting twist. Alongside the open Llama 5, Meta just launched Muse Spark 1.1 as a paid, proprietary coding model at $1.25 per million input tokens and $4.25 per million output tokens. That's Anthropic and OpenAI pricing territory. And Muse Spark also powers the new consumer Meta AI assistant that plans your kitchen renovation, books your calendar, and generates slides autonomously. So Meta is running a barbell strategy: open weights for the general model to seed the ecosystem and starve competitors of oxygen, closed and paid for the frontier coding and agent stack where they actually want to make money. It's the same playbook Google tried with Android, and it's why they signed that open-weight letter last week. Open weights aren't charity. They're a moat against OpenAI and Anthropic charging rent on API access. The one weak spot: Llama 5 still trails on coding benchmarks. DeepSeek and the closed frontier models beat it there. Which is probably why Muse Spark exists.
OpenAI quietly launched something last week that I think matters more than the market has priced in. It's called Presence, and it's not a model. It's an enterprise product for deploying real AI agents into call centers, billing departments, insurance claims, IT desks, sales development. The pitch: task-specific agents with company-defined policies, guardrails, simulations, human escalation, and a Codex-driven improvement loop that watches production traffic and proposes updates. The proof point is OpenAI's own English phone support line, 1-888-GPT-0090. Presence handles 75% of inbound issues without a human, and the Codex loop cut human handoffs by 15 percentage points in 10 days. Early customers include BBVA Mexico, SoftBank in Japan, and Retail Insurance Australia. It's not self-serve. Deployments are led by OpenAI Forward Deployed Engineers and a small set of systems integrators, which tells you exactly who the customer is: F500 companies with big contact centers and compliance teams. Two things worth noting. First, this is OpenAI moving from selling model access to selling end-to-end enterprise software. Stickier revenue, bigger contracts, but they'll now compete with some of their own API customers, the Sierras and Deciagons of the world. Second, look at where this converges with Coinbase's pitch. Brian Armstrong is out this week saying Base has processed over 100 million AI payments via x402 and USDC, framing agentic finance as the next adoption curve. Agents that resolve billing tickets, book travel, and pay for compute need payment rails. Which is the perfect setup for the next segment.
Lightning Labs shipped Wavelength last week, and if you care about Bitcoin as money rather than as a treasury asset, this is the story of the month. Wavelength is a self-custodial SDK that lets a developer or an AI agent add Bitcoin payments to an app with a few API calls. No running a node. No managing channels. No sourcing liquidity. Keys stay on device. Users can unilaterally exit to on-chain Bitcoin whenever they want. Under the hood it supports three rails: on-chain Bitcoin, Lightning via atomic swaps through Loop, and an Ark-style off-chain settlement layer that batches transfers. Every off-chain payment still uses a standard BOLT 11 invoice, so it interoperates with the existing Lightning network. Fees during the closed alpha are 1 basis point. The killer feature is the agent integration. Wavelength exposes its API through the Model Context Protocol, so an AI agent can hold a Bitcoin balance and pay for compute, data feeds, or API calls in fractions of a cent. It also plugs into L402, the HTTP 402 payment-required standard, so APIs can demand small Lightning payments automatically, no accounts, no keys, no Stripe. Think about the stack that's now assembled. OpenAI Presence deploys autonomous agents into enterprises. Coinbase pushes x402 and USDC on Base for agent payments. Lightning Labs offers the Bitcoin-native equivalent with self-custody and true micropayment economics. The x402 side, by the way, just had a rough week. Researchers disclosed 31 vulnerabilities affecting 99% of x402 payment operators, including validated free-shopping exploits. Self-custodial Bitcoin with unilateral exit looks pretty good next to that. Currently Signet and testnet, mainnet is invite-only. But the direction is clear: agents will need to pay for things, and Bitcoin is quietly building the rails.
The GENIUS Act was signed a year ago, and the regulatory picture is finally coming into focus, and it's not what U.S. issuers hoped for. The one-year deadline for finalized rules just slipped, but the shape of the framework is clear. Two developments this week make the split explicit. First, the FDIC published proposed reporting forms for permitted payment stablecoin issuers. Weekly confidential filings on reserves. Quarterly reports mirroring bank Call Reports. Separate forms per brand. Full bank-style supervision. Second, and more consequential, Circle received final OCC approval to form Circle National Trust, the first federally chartered trust bank for a digital asset company. USDC now sits inside a nationally chartered bank with about $73 billion in circulation. That's the good news for Circle. Here's the bad news. The GENIUS Act flatly bans permitted stablecoin issuers from paying any yield to holders. No interest, no rewards, no token distributions. So Circle, now the most regulated stablecoin issuer on earth, cannot share reserve income with USDC holders. Tether? USDT is offshore. Not subject to the Act. Tether launched USA-T as a compliant U.S. product while keeping USDT outside the perimeter, meaning globally they can still capture the yield game. The competitive asymmetry is stark. Offshore stablecoins and DeFi lending can pay mid-single-digit yields. U.S. regulated USDC cannot. Circle's business model, which relied heavily on reserve yield through 2022 to 2024, now has a ceiling on how much it can share with holders to grow supply. There's a bipartisan PARITY Act draft that would fix the tax treatment for everyday stablecoin use, but yield restrictions aren't going away. The result: regulated dollar rails inside the U.S., yield-bearing dollar rails everywhere else. Tether just got handed a durable global moat by U.S. regulation. And Circle's response, buying nearly 1,000 IBM blockchain patents this week, looks less like offense and more like building defensive walls around the regulated piece of the market they're allowed to own.
A prediction: within 18 months, the phrase enterprise AI agent will mean an autonomous process that pays for its own compute, data, and services with a stablecoin or with Bitcoin. The infrastructure went from vaporware to shippable in under a year. The winners won't be the flashiest models. They'll be whoever owns the rails those agents use to move money.