Gm, and happy Thursday!

NVDA beats earnings while crypto indices rally even harder. The bid from treasury companies was sidelined for this rally; could these vehicles be the next big buyer?  

Below, we explore Darkbloom, a side project out of Eigen Labs offering Apple Silicon-native AI inference from a distributed network of contributors.

NVDA’s earnings beat after hours Wednesday boosted what was a soft equities market over the past week as the chip leader gaps up 6% into Thursday’s open. The Nasdaq 100 opens up 1%, with longer-tail AI names rallying even harder on the back of this. The beat reinvigorates the AI trade that lost its fervor over the past several weeks, keeping risk in play in equities and maintaining a constructive backdrop for crypto, which now leads the board. Over the past week, every crypto index we track registered strong positive returns, except for crypto miners. Breadth and strength are back in crypto, with the Lending sector leading this rally. Blue chips in the sector like AAVE, KMNO, and SPK are all up over 30% in this rally. 

Importantly, treasury companies are nowhere to be found in this rally, and the ATM printer does not account for any of the bid on the majors. The bid from these vehicles was largely sidelined throughout July and August, with net purchases remaining paltry and BTC flows actually negative $100M in both months. 

In liquid markets, there’s alpha in knowing who the next buyer is. Looking at aggregate mNAVs for treasury companies across BTC, ETH, and SOL, the multiples have set a modest floor and are trending back up towards 1. Should these multiples flip to a premium, the ATM facilities are back in play, allowing these vehicles to sell shares for spot and resume accumulation. A flip above 1 on these metrics over the coming weeks or months could accelerate what has already been a strong market. Until and unless this happens, the ATM facility, which accounted for a majority of the financing for treasury company purchases, remains off the table. 

Regardless, the backdrop in risk assets remains constructive, ETF holdings are rising, and the trend is now up, setting up a favorable outlook for crypto into year-end. 

— Luke

Darkbloom's no-token DePIN test

DePIN teams spent the last cycle arguing that token emissions are the price of bootstrapping distributed hardware supply. A side project out of Eigen Labs quadrupled its machine count over a single weekend without one.

Darkbloom, a distributed inference network that routes AI jobs to idle Apple Silicon Macs, crossed 1,000 machines on Monday, per the project, up from 250 on Friday. The network began as an engineer's vacation experiment in April, hit #1 on Hacker News within two days, and went live on OpenRouter in June. Per a timeline from Eigen's Pratik Gandhi, throughput grew from 100M tokens per day in mid-June to ~10B today. On August 20th the network moved from OpenRouter's free tier to paid status, and Eigen engineer Gajesh Naik cited $102K in ARR.

Mac owners install a provider client, keep the machine online, and earn from inference jobs matched by a coordinator. Developers point an OpenAI-compatible endpoint at the network and pay rates set around 50% below list prices for comparable open-weight models, on the logic that the hardware is paid for and the marginal cost is electricity. Privacy runs through Apple's stack, with requests encrypted end-to-end, provider keys generated inside the Secure Enclave, attestation traced to Apple's root certificate authority, and hardened runtime blocking operators from inspecting the inference process on their own machines. 

Apple Silicon has spent two years becoming the default substrate for local inference, on the strength of unified memory that delivers two to three times the bandwidth per dollar as Nvidia's desktop hardware. Exo, the open-source project that pools multiple Macs into a single inference cluster, has drawn more than 43K GitHub stars, and newer builds shard models across machines over Thunderbolt 5 RDMA. Apple's own Private Cloud Compute established the attestation playbook Darkbloom borrows. 

The earnings claims deserve scrutiny, and the team's own updates now do some of that work. Naik cites $120 to $200 per month per machine on average, and Darkbloom's original calculator projected up to $4,983 per year. The rebuilt calculator shipped Tuesday with a duty cycle slider, and the defaults are much lower: a Mac Mini M4 Pro shows $170 per month at 100% duty cycle, but the slider defaults to 5%, which works out to roughly $8.50.

The team's disclosure notes "significant variation in earning levels among providers using the same machine type" and says the 5% default reflects that. Set the headline claims against the disclosed $102K ARR spread across the 250 machines online last week and run-rate revenue works out to ~$34 per machine per month, closer to the 5% default on a high-end Mac Studio ($37.50). The team acknowledged the gap over the weekend, writing that the calculator numbers show "what's possible," and the new defaults land near where the aggregate math already pointed. 

Supply then outran the network's ability to use it. In a Wednesday update, the team said capacity grew more than 10x over the weekend against a tripling in daily volume, and it paused new enrollment for Macs below 48GB of memory because its routing cannot yet tell "a healthy small Mac from an unreliable one," leaving many new providers idle with few requests. The same post claims some sub-48GB machines earn close to $1,000 a year, per the team, which reinforces the concentration story. A network where a few well-placed machines earn four figures while hundreds sit idle produces exactly the variance the calculator's 5% default now admits.

Darkbloom lands as an argument against DePIN's core assumption. Render, io.net and Akash spent years and heavy token emissions recruiting GPU supply, while Darkbloom pulled 750 machines onto the network in a weekend with fiat payouts and an earnings calculator, and the OpenRouter throughput comps sharpen the point. Per OpenRouter's provider pages, Darkbloom cleared a record ~11B tokens on Wednesday across three models. That run rate sits above io.net, whose daily totals have ranged 2-6B tokens since late May across five models, and inside the band of Chutes, the Bittensor subnet whose recent days swing between 5-17B. AkashML leads the cohort, with recent days between 18-36B. Each of those networks sits atop a token incentive layer. 

Darkbloom reached the middle of the pack in ten weeks on fiat payouts and a base reward pool. Provider retention remains untested, and the first sustained demand drought will test whether 1,300 Macs stay online for single-digit monthly payouts. Demand decided every DePIN winner to date, and Darkbloom's midweek fixes all target that constraint: Qwen 3.6 uptime back above 99%, GPT-OSS-20B live as a second paid model, and more models queued because, in the team's words, adding a model is the fastest way to pull in demand. The supply side of the experiment ended with new enrollment paused after supply outpaced demand. The demand side starts now.

— Nick

The Bulls Are Back: Treasury Intervention, SEC Crypto Rules & Trump Says Hyperliquid

The 0xResearch crew examined the Treasury's buyback expansion from $2B to $4B and a reported plan under consideration for up to $1T in 30-year purchases, framing the move as forward guidance that removes duration from the market rather than QE. Mark argued that AI corporate bond issuance crowded out both Treasurys and crypto flows over the past 12 months, keeping long-end rates elevated. The bulk of the episode centered on the SEC's Regulation Crypto Assets proposal, covering the $5M startup exemption, tiered fundraising exemptions up to $75M, the Form TR off-ramp's "sufficiently functional" standard, and airdrop language that treats subsequent recipient action as consideration. The panel closed on Trump's Hyperliquid name drop, CFTC perp onshoring efforts, and spot ETF inflows posting their largest uptick in 10 weeks.

Kalshi’s 15-Minute Markets Now Drive Bitcoin Price Discovery

This research report by Synth Data examines how Kalshi’s rapidly growing 15-minute Bitcoin prediction markets are increasingly functioning as a source of real-time BTC price discovery. It argues that the market’s high-frequency contracts can incorporate information and express short-term expectations faster than traditional spot markets, creating a feedback loop between prediction-market prices and Bitcoin’s underlying price. The research highlights the growing importance of prediction markets as financial infrastructure, particularly as liquidity and participation deepen. More broadly, it suggests that Kalshi’s evolution from a venue for event contracts toward a continuous market for short-horizon asset prices could have meaningful implications for how crypto markets discover prices.