Bitcoin closed the worst first half in years, dropping into the high 50s after a brutal June. Spot Bitcoin ETFs bled a record 4.5 billion dollars for the month — the worst on record. Meanwhile, BNY Mellon added USDC minting and burning directly inside its institutional custody platform, DeepSeek dropped an open source inference framework that speeds up serving by up to 85%, and BMW is putting Figure's next-generation humanoid onto its Spartanburg factory floor. Four stories worth your time today.
Let's start with the number that matters. Bitcoin fell in both Q1 and Q2 of 2026. That has only happened three times in Bitcoin's history — 2018, 2022, and now. In both prior cases, the second half brought no rescue. So the historical base rate here is ugly.
June specifically was a bloodbath. US spot Bitcoin ETFs lost 4.5 billion dollars, beating the previous worst month by 29%. Nine consecutive days of redemptions closed out the period. Year to date, ETF outflows are now 5.5 billion. That's the same instrument that was supposed to be the permanent institutional bid.
Citi cut its 12-month price targets on both Bitcoin and Ether, scrapping its ETF inflow forecasts entirely, citing stalled US crypto legislation and weak demand. Options traders are loading up on 50,000 dollar puts. Gold futures just flashed a death cross. Bitcoin bounced to 60,000 after Warsh comments this morning, but the setup is fragile.
Here's the more interesting read though. Open interest collapsed after 8.35 billion in long liquidations during the Q2 flush. Leverage is largely gone. Talos notes Q3 opens with thinner liquidity but a much cleaner positioning picture. Some analysts think a drop another 5,000 dollars lower — into the low 50s — would put Bitcoin within 10% of its realized price, which has historically marked bear market bottoms.
Strategy, formerly MicroStrategy, is even talking about buybacks of MSTR and STRC and expanded cash reserves, with potential Bitcoin sales on the table to defuse death spiral fears. That's a meaningful shift in posture from a company that spent three years buying every dip with leverage.
The uncomfortable truth: the ETF-era floor thesis is being tested in real time. If passive institutional flows don't come back in Q3, the buyer of last resort question gets very sharp, very quickly.
While retail flees, the plumbing keeps getting built. BNY Mellon — which oversees roughly 59.3 trillion dollars in assets under custody — just added Circle's USDC to its Digital Asset Custody platform. Institutional clients can now store, transfer, mint, and redeem USDC directly through the bank. It's the first stablecoin on the platform, and BNY says more issuers are coming.
This is bigger than it sounds. BNY has been Circle's primary reserve custodian since 2022. Now they're offering the full lifecycle — client walks in with dollars, walks out with USDC, or vice versa, without ever leaving the regulated custody rail. Mint and burn inside the bank. That's the on-ramp and off-ramp that institutional treasurers have been asking for.
Credit Agricole's CACEIS launched a euro-pegged stablecoin, EURXT, on Ethereum today. 20 million tokens at launch, backed one-to-one by euro reserves. That's a Tier 1 European bank issuing on a public chain. Bank of Korea's governor spent his time at the ECB Forum talking about tokenized government bonds and a unified ledger plan.
And on the private consortium side, Visa, Mastercard, Coinbase, and others are backing Open USD — a partner-led stablecoin where reserve earnings get shared with partners rather than pocketed entirely by the issuer. That's a direct shot at Circle and Tether's economics.
So while Bitcoin ETF flows collapse and prices sag, the actual TradFi integration keeps compounding. Stablecoin rails, tokenized bonds, bank-issued digital cash. The macro price action and the infrastructure buildout are telling completely different stories right now — and the infrastructure story is the one with a five-year lead time. It's arriving on schedule whether the market is green or red.
Switching to AI. DeepSeek just open sourced DSpark under MIT license — a serving framework that speeds up LLM inference by 60 to 85% per user, and 51 to 52% at aggregate throughput. It's not a new model. It's a smarter way to run the models you already have.
The core trick is speculative decoding. A small, fast draft model proposes candidate tokens in parallel. The big target model then verifies them in batches, rather than generating one token at a time. DeepSeek's twist adds a semi-autoregressive component so draft tokens stay coherent with each other, plus a confidence-scheduled verifier that dynamically decides how much compute to spend based on how likely each draft is to survive. Hardware-aware, load-aware, and it preserves the target model's output distribution exactly.
They tested it against Qwen3 and Google's Gemma as target models, not just their own V4-Pro. In production traffic, average accepted token length went up 26 to 30% versus the previous best open framework. Everything — the framework, the paper, and DeepSeek V4-Pro itself — is on Hugging Face and GitHub.
And V4 proper launches mid-July with a 1-million-token context window across the whole lineup, plus peak and off-peak API pricing. Peak hours cost double.
The strategic read here matters. US export controls have squeezed China's access to Nvidia's top chips. DeepSeek's response is to squeeze more tokens out of every chip they can get. DSpark shifts the Pareto frontier — higher throughput with fewer high-end GPUs. That's a direct answer to a hardware constraint via software optimization, and they're giving it away.
Meanwhile Qwen 3.6-27B dropped as an open-weight model that runs locally on a consumer gaming PC — 16.8 gigabytes quantized, roughly 50 tokens per second on a 32-gig GPU, scoring 77.2 on SWE-bench Verified. That's frontier-adjacent coding performance without a cloud API or subscription.
The open weight ecosystem out of China is not catching up to the frontier — on serving efficiency and local deployment, it's setting the pace.
Last one. Humanoid robots stopped being a demo reel this quarter.
BMW is deploying Figure AI's Figure 03 at its Spartanburg plant in South Carolina — this after Figure 02 already proved out in the body shop supporting welding and sheet metal handling across 30,000-plus X3 units. The new deployment moves upstream into logistics sequencing: Figure 03 picks unsorted components from bins, sorts them into trolleys, and an automated tugger delivers them just-in-sequence to human assemblers. Figure 03 has tactile hands with palm cameras, wireless charging for uptime, and speech-to-speech audio.
Over in China, AGIBOT ran a six-day livestream of its G2 humanoid on Longcheer's active mass production line in Nanchang. 64 hours of continuous operation. 64,828 tasks completed. 99.99% task success rate. And they announced their 15,000th unit delivered — production capacity went from 1,000 to 5,000 units in a year, then 5,000 to 10,000 in three months. AGIBOT claims 39% of global humanoid shipments.
CATL — the world's largest battery maker — put Galbot's S1 humanoid on its battery pack assembly lines. Dual arms, 50 kilogram payload, 8 hours of continuous operation, powered by CATL's own batteries. The two companies signed a global strategic cooperation agreement. Galbot's co-founder predicts full commercial factory rollout within two years.
And Agility Robotics' Digit just hit one full year of continuous paid operation at GXO — plus deployments at Amazon, Toyota, and Schaeffler. This is not pilot revenue anymore. A rumored 2.5 billion dollar SPAC deal is on the table, with production scaling past 10,000 units per year at their Salem, Oregon facility.
Zoom out. Six months ago the humanoid conversation was still about whether these things could reliably grab a tote. Today we have multi-year production deployments at BMW, Amazon, Toyota, GXO, and CATL, plus a Chinese vendor shipping 15,000 units. The bottleneck is no longer the robot. It's manufacturing capacity, safety certification, and integration engineering. Physical AI is now a supply chain problem, not a research problem.
One prediction to close on. If Bitcoin does grind into the low 50s this quarter, the people buying will not be the ETF holders who just sold. They'll be the same institutions that just quietly plugged USDC into BNY's custody rail. The infrastructure is being built to buy — the question is just at what price.