Bitcoin is grinding against $80,000 again — the third test of the year — while April closed as the strongest month in twelve. ETF inflows are roaring back, with $629 million landing on May 1 alone. Lightspark just made one of the biggest payments deals in Bitcoin's history, partnering with Visa to issue cards across more than 100 countries. Xiaomi quietly dropped a trillion-parameter open-weight model that codes for eleven hours straight. And Brazil's central bank slammed the door on stablecoin settlement for cross-border payments. Let's get into it.
David Marcus is back, and this time he's not running a Facebook coin. Lightspark announced at Bitcoin 2026 that it's becoming a Principal Visa Member in Europe and rolling out crypto-backed debit cards across more than 100 countries, accepted at over 175 million Visa merchants.
The product is called Grid Global Accounts. It's an API platform — not a consumer app — that lets fintechs, banks, and businesses issue Visa cards funded by Bitcoin, stablecoins, or fiat. Bitcoin payments route through Spark and the Lightning Network. Stablecoins ride on Solana, Base, and Spark. And Lightspark handles the compliance and settlement plumbing in the background.
The demo Marcus showed is the kind of thing that used to take a fintech five years to build. A creator in Mexico receives $5,000 from a U.S. platform, swipes a Visa card globally, then sends funds to a Brazilian Pix account in seconds — all from one wallet holding Bitcoin, dollars, and stablecoins simultaneously. There's even a delegated AI agent feature where you can hand a scoped, limited-use card to an agent to make purchases on your behalf.
Here's what's interesting from a Bitcoin-maximalist lens. Marcus is explicit that stablecoins are the transport layer, not Bitcoin. The pitch is rail-agnostic — "a dollar is a dollar on any network." Bitcoin sits underneath as one option among many, with Spark giving it the speed needed to compete with stablecoin rails. So this is Lightspark moving from a Bitcoin company to a payments company. Some people will hate that. But the practical effect is that Lightning and Spark just got pushed onto 175 million merchant terminals, and the same infrastructure that settles a USDC payment can settle a Bitcoin one. The plan is 75 countries and 100 Visa markets by the end of 2026.
Meanwhile, Brazil's central bank went the other direction this week and banned stablecoin and crypto settlement in cross-border payments for fintechs and payment firms. Individuals can still hold crypto, but the back-end rail is closed. Two stories, opposite directions, same week — that's where we are with payments right now.
Bitcoin is sitting around $78,000 heading into a heavy macro week. April closed up roughly 13% — the best monthly performance in a year — but BTC is still testing the same $78,000 to $80,000 wall it's hit twice already in 2026.
The setup has two competing stories. On one side: ETF inflows are real and accelerating. May 1 alone brought in $629 million, the largest single-day inflow since mid-April, with BlackRock's IBIT pulling $284 million of that. Total ETF assets are now around $103 billion, about 6.66% of Bitcoin's market cap. Whales — wallets holding 1,000 to 10,000 BTC — have added roughly 240,000 BTC since December. Institutions added another 92,900 BTC last month per Bitwise.
On the other side: CryptoQuant is flagging that April's rally was futures-driven, not spot-driven. Apparent spot demand was actually negative even as futures demand climbed. That's the same pattern they saw at the start of the 2022 bear market. Negative funding rates on perps suggest retail isn't buying this either.
So the bull case is simple — clean break above $80,000 and we're looking at $86,000 to $88,000, with the 50-week moving average around $93,000 to $95,000 as the next major level. The bear case is uglier — there's a head and shoulders pattern with a measured move down to $48,000 if support around $77,000 cracks.
The wildcard is the Fed. Markets are pricing a 99% probability of a rate hold this week, but Powell's tone matters more than the decision. There's also the leadership transition to Kevin Warsh on May 15, which nobody is really pricing yet. And oil above $100 a barrel is keeping inflation sticky, which makes near-term cuts unlikely.
Mining stocks, by the way, are telling their own story. Every major miner is up in 2026, some by as much as 85%, while Bitcoin itself is still down on the year. That divergence usually closes one way or the other.
Three big AI releases this week, and the through-line is the same: inference is getting dramatically cheaper, and the open-weight side is no longer playing catch-up.
Xiaomi released MiMo-V2.5-Pro under the MIT license. It's a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per request. Context window: 1 million tokens. In Xiaomi's internal tests, it built a full compiler in 4.3 hours across 672 tool calls and passed all 233 hidden tests. In another demo, it autonomously wrote a desktop video editor — about 8,000 lines of code — in 11.5 hours and 1,870 tool calls. Xiaomi claims it uses 40 to 60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, or GPT-5.4 to reach similar results. Trillion-parameter MIT-licensed weights, sitting on Hugging Face, free for commercial use.
Mistral countered with Medium 3.5 — a dense 128-billion-parameter model that unifies chat, reasoning, and code in one weight set. The clever bit is a per-request reasoning_effort toggle, so you only pay for heavy thinking when you actually need it. 256,000-token context, runs on as few as four GPUs, $1.50 per million input tokens, $7.50 per million output. They're explicitly positioning it as the only non-Chinese open-weight flagship in this tier — competing with Qwen, Zhipu GLM, and Xiaomi MiMo.
And on the hardware side, Moreh published numbers showing Tenstorrent Galaxy nodes running DeepSeek R1 671B at 1.68 times the throughput of an NVIDIA DGX A100 baseline, with faster time-to-first-token — at roughly one-third the cost. Galaxy nodes run about $110,000 versus $330,000 to $550,000 for DGX systems. If those numbers hold up under independent testing, the economics of running production inference shift meaningfully. You don't have to be hyperscale to afford serious capacity anymore.
Google also dropped TurboQuant, a KV-cache compression technique that cuts memory use up to 6x during real-time chat without quality loss. Same direction: longer contexts, more reasoning, less hardware.
The pattern across all of this is brutal for proprietary frontier labs. Open weights are now competitive on coding benchmarks, hardware alternatives are real, and memory optimizations are killing the moat that came from owning massive GPU clusters. The cost-to-serve curve for capable AI is collapsing fast.
A few funding rounds worth tracking, plus a notable foundation launch.
Squads raised $18 million led by Solana Ventures, with Coinbase Ventures and Haun participating. Their product, Altitude, is a financial operating system built entirely on stablecoin rails — treasury stays in stablecoins, settlement is instant, and the company never holds customer funds. They've processed over $200 million in payments since launching in December. The market context here matters: Stripe acquired Bridge, Mastercard acquired BVNK. Stablecoin business finance is consolidating fast.
Liquid closed an $18 million seed led by Neo and Left Lane Capital, with Paradigm and General Catalyst returning. The pitch is one interface for crypto, equities, FX, commodities, and pre-IPO assets, with up to 200x leverage and 24/7 markets. Since launching in August 2025, they've done over $3 billion in volume across roughly 40,000 users. The thesis is that retail traders want professional tools and continuous markets — and they don't care about the asset class boundaries that traditional brokers enforce.
Nuva, an RWA yield platform co-incubated by Animoca Brands, raised $5.2 million led by Morgan Creek Digital. Users deposit USDC into vaults backed by tokenized real-world assets — initially using on-chain RWA from Figure Technologies — and receive vault tokens whose price reflects underlying yield. This is part of a broader theme: Mike Cagney's Figure had a billion-dollar month bringing real-world assets, lending, and equities onchain.
And MARA Holdings launched the MARA Foundation at Bitcoin 2026 with a $100,000 initial fund and three candidate projects up for community vote: 256 Foundation for open-source mining, Libreria de Satoshi for Latin American Bitcoin education, and SafeNet for community wireless networks. The stated focus is hardening Bitcoin against future quantum threats and expanding self-custody access in underserved regions. $100,000 is small money for MARA, but the signaling matters — a major public miner formally putting weight behind protocol resilience and grassroots adoption.
The pattern across all four: stablecoin infrastructure, tokenized real-world yield, multi-asset trading, and Bitcoin protocol resilience. Notably absent from this week's rounds — anything pure DeFi or pure speculation. Capital is going toward the boring, unsexy plumbing of finance.
One observation to close on. A new CoinDesk poll out today shows U.S. voters rank crypto dead last among their political priorities, and most still trust banks more than crypto for financial access. Meanwhile, Lightspark just plugged Bitcoin and stablecoins into 175 million Visa terminals without anyone needing to know or care. The industry's biggest wins from here probably won't come from convincing the public to love crypto. They'll come from making the rails invisible.