Bitcoin slipped under $80,000 this morning after hitting resistance at $82,800, but spot ETFs pulled in $1.105 billion this week, the strongest inflow since January. Tom Lee says a May close above $76,000 confirms a new bull market. Over at Consensus Miami, Bitcoin treasury firms outlined a $3 trillion opportunity in BTC-backed digital credit, while Strategy CEO Phong Le admitted the company will sell Bitcoin when it suits them, breaking the never-sell mantra. Mistral dropped a 128 billion parameter open model. Amazon, Coinbase, and Stripe just wired up payment rails so AI agents can spend stablecoins. And the CLARITY Act is finally heading for Senate Banking markup next week. Let's get into it.
The CLARITY Act has been stuck in Senate purgatory for months, but it looks like the dam is breaking. Senator Tim Scott, who chairs Senate Banking, says the committee is targeting a bipartisan markup the week of May 11. The White House wants this on the president's desk by July 4. That's roughly 58 days from now. The bill already cleared the House 294 to 134 and got through the Senate Agriculture Committee back in January. What it does is split jurisdiction between the SEC and CFTC for digital assets, finally giving the U.S. a market structure framework. The big sticking point has been stablecoin yield. Banks were terrified that yield-bearing stablecoins would drain deposits out of the traditional system. Senators Tillis and Alsobrooks brokered a compromise that draws a bright line. Passive yields that look like deposit interest are banned. But activity-based rewards tied to actual usage, payments, transfers, trading, those are allowed. The reframe is important. Stablecoins become payment rails, not savings accounts. Coinbase and Circle are on board. Banks got what they wanted on deposit competition. But the banking lobby is still fighting hard to slow this down, and Senator John Kennedy remains a no vote on the Republican side. There's also a new wrinkle around DeFi liability protection that's drawing law-enforcement objections. The SEC has scheduled a CLARITY Act roundtable for May with CFTC officials to hammer out which assets sit where. This is meant to turn the March 17 joint taxonomy, which classified 16 assets as commodities, into actual federal law. Cynthia Lummis and Bernie Moreno are warning that if Congress misses this window, the next real shot might not come until 2030. Five years of regulatory limbo. So watch next week. If markup happens and survives, floor action in June is plausible, and the July 4 signing ceremony stops being a fantasy.
The biggest structural shift in AI right now isn't model benchmarks. It's that AI agents are getting bank accounts. Today Amazon, Coinbase, and Stripe announced payment infrastructure that lets AI agents make autonomous USDC payments. Coinbase's x402 protocol plugs into Amazon Bedrock AgentCore so agents can buy APIs, web content, and online services on their own. Amazon says future versions will handle hotel bookings and merchant payments. Tether-backed Oobit went further. They launched Visa-backed Agent Cards where each card is bound to a single AI agent and pulls USDT directly from Tether's treasury, no fiat conversion. The cards work with OpenAI, Claude, AutoGen, and LangChain, with spend limits enforced at the transaction level after KYB checks. Use cases are unglamorous but real. Subscription renewals, ad spend, cloud bills, all running on autopilot with audit trails. CZ floated reviving Binance.US specifically because he thinks BNB Chain is the optimal rail for agent-to-agent payments. At Consensus, executives from Bridge and Deus X Capital said AI agents using blockchain rails will lead the next stablecoin boom alongside corporate treasury flows. eToro's Yoni Assia said DeFi isn't dead, it's going mainstream because agents need permissionless rails to operate. On the enterprise side, the buildout is moving fast. NVIDIA and ServiceNow announced Project Arc, a long-running autonomous desktop agent for knowledge workers, running in sandboxed environments via NVIDIA OpenShell. Google rebranded Vertex AI as the Gemini Enterprise Agent Platform with a $750 million partner fund. Citi launched its own internal Arc platform to deploy governed agents across banking workflows, starting with wealth management prep work. The pattern is consistent. Agents are no longer chatbots. They're economic actors with credentials, budgets, and audit logs. The infrastructure layer being built right now, payment rails, governance, sandboxes, is what determines who captures the value when this scales.
Two big open-weight model drops this week worth paying attention to. Mistral released Medium 3.5, a dense 128 billion parameter model with a 256k context window. It's multimodal, handles vision, and ships with configurable reasoning effort per request, so you can dial it from quick reply to long agent session without swapping models. They published weights on Hugging Face under a modified MIT license. Self-hosting works on four GPUs. Pricing is $1.50 per million input tokens and $7.50 per million output tokens. On SWE-Bench Verified it scored 77.6%, which is competitive with much larger models for real coding work. Mistral also rolled out Vibe, their coding platform with cloud-based remote agents. You can teleport a local CLI session to the cloud, keep history and approvals, run isolated sandboxes, and have the agent open a GitHub pull request when it's done. Integrations with Linear, Jira, Sentry, Slack. Mistral is leaning hard into European data sovereignty, including an $830 million loan for a Paris-area data center. Meanwhile, the U.S. AI Safety Institute, CAISI, published its evaluation of DeepSeek V4 Pro. The headline finding is that V4 Pro is the strongest Chinese model they've tested, beating other PRC models across cyber, software engineering, natural sciences, abstract reasoning, and math. But it lags frontier U.S. models by about 8 months. CAISI puts it at roughly GPT-5 level capability. DeepSeek's own self-reported numbers overstate where it actually lands. On cost efficiency it's mixed, ranging from 53% cheaper to 41% more expensive than the U.S. reference model depending on benchmark. The V4 release also includes a flash variant at 284 billion total parameters, both built on Mixture-of-Experts, with a default 1 million token context window. Long context plus open weights plus reasonable cost is the recipe that makes agentic workflows actually viable. The 8-month gap is real, but the gap is no longer two years.
Something shifted on Strategy's earnings call Monday. CEO Phong Le said plainly that the company will sell Bitcoin when it's advantageous. Michael Saylor added that Strategy would probably sell some Bitcoin to fund a dividend, in his words, to inoculate the market. This is a meaningful departure from years of never-sell rhetoric. Strategy holds 818,334 BTC as of May 3, up 22% year-to-date. Samson Mow defended the move publicly, but the optics matter. The largest corporate Bitcoin holder is now openly contemplating monetization. Price action reflects the shift. Bitcoin tested $82,800 as resistance and dropped under $80,000. Analysts say reclaiming $88,000 as support and a slowdown in profit-taking are the prerequisites to confirm a cycle bottom. Tom Lee at Consensus said closing May above $76,000 would confirm a new bull market driven by tokenization and AI agentic finance. VanEck's Matthew Sigel laid out a $1 million Bitcoin target within five years, comparing adoption to the gaming industry's mainstream shift. On the institutional side the picture is genuinely strong. CoinShares reports fund managers are increasing Bitcoin exposure as sentiment rebounds. BlackRock's IBIT alone holds $66.7 billion. Cumulative net inflows into U.S. spot Bitcoin ETFs sit around $59.7 billion. Morgan Stanley and Charles Schwab are pushing direct crypto trading into ordinary brokerage accounts because they see the demand in their own client base. BNY is launching institutional Bitcoin and Ethereum custody for UAE clients out of the Abu Dhabi Global Market. On the mining side, the squeeze is brutal. Core Scientific posted a $347 million quarterly loss. They mined 279 BTC, down 45% year-over-year, and AI hosting is now their top revenue line, not Bitcoin mining. Trump-linked American Bitcoin missed Q1 estimates with an $82 million loss. The miner business model is bifurcating fast. Either you have AI hosting deals or you're getting crushed. One more story worth flagging. A California man, Marlon Ferro, got 78 months in prison for a $250 million crypto theft conspiracy that involved breaking into victims' homes to physically steal hardware wallets when remote hacking failed. Self-custody matters. So does operational security.
Strategy saying the quiet part out loud, that they'll sell Bitcoin when it suits them, is the most honest thing any treasury company has said in two years. The never-sell narrative was always marketing. Treat it accordingly.