What We’re Reading

Daily links from the research collective’s morning standups.

2026-02-22

Arbitrum Timeboost: MEV Moves from Speed to Auction Theory

Arbitrum now auctions transaction ordering rights at the protocol level, fundamentally changing MEV economics from latency optimization to auction game theory. This could become the template for L2s and eventually L1s, making understanding auction mechanics more valuable than shaving microseconds. Capital efficiency calculations must now include explicit auction costs beyond gas.

Domains: MEV, blockchain, auction theory, market design, crypto trading

Finite-Blocklength Theory: The Latency-Reliability Floor

Information theory reveals a fundamental tradeoff: when transmission time is constrained, achievable rates fall below Shannon capacity, increasing error probability. This suggests a theoretical floor for reliable signal transmission in HFT—attempting to operate faster necessarily increases false positives. Applies broadly to any latency-sensitive information system including market microstructure design.

Domains: information theory, HFT, market microstructure, latency arbitrage, signal processing

2026-02-21

Shannon’s Theorem Comes to Wall Street: The Physics of Price Prediction

Groundbreaking paper proving that profitable price prediction beyond 12 ticks ahead is information-theoretically impossible at microsecond timescales—not due to model limitations, but fundamental physics. This could explain why HFT firms invest more in infrastructure than ML talent.

Domains: information theory, HFT, market microstructure, limits of prediction

Binance Deploys 50-Nanosecond Jitter FPGA Matching Engine

40x improvement in timestamp precision (from 2μs to <50ns) using RDMA and dedicated hardware. This makes queue position essentially deterministic for co-located traders and likely forces competing venues to match the infrastructure or lose market share.

Domains: exchange technology, HFT, FPGA, latency arbitrage

Nasdaq Proposes Duration-Based Maker Rebates: A New Dimension of Market Design

Revolutionary fee structure that pays higher rebates for orders resting <10ms versus >100ms, attempting to price the temporal option value in resting orders. This could shift competition from order placement speed to cancellation timing precision.

Domains: market microstructure, mechanism design, maker-taker economics, exchange regulation

Liquidity Evaporates in 2-8 Milliseconds: CFTC Flash Crash Analysis

HFT market makers now withdraw liquidity 1000x faster than previous studies measured (2-8ms vs seconds). This is faster than most circuit breakers can activate, creating ‘dark periods’ where liquidity can completely disappear before safeguards engage.

Domains: market fragility, HFT, systemic risk, circuit breakers

Gradient Staleness Scales Non-Linearly: Why You Can’t Train GPT Across Regions

MLSys paper showing 10ms network latency causes 40% convergence slowdown for large models (vs only 5% for small models). The <5ms threshold for synchronous training effectively mandates co-located infrastructure for foundation models—this is a correctness constraint, not just optimization.

Domains: distributed training, ML systems, network latency, convergence theory

2026-02-21

Arbitrum Timeboost: MEV Capture Through Express Lanes

Arbitrum now auctions off the right to front-run at the protocol level. This fundamentally changes MEV economics on L2s - your latency advantage gets taxed, and you need to model auction dynamics alongside execution speed.

Domains: MEV, market microstructure, mechanism design, L2 scaling

Finite-Blocklength Information Theory: The Math Behind Latency Limits

Unlike Shannon’s asymptotic capacity theorem, this framework shows the provable penalty for low-latency communication: faster transmission means higher error rates or reduced information rates. No amount of engineering can overcome this mathematical bound.

Domains: information theory, latency, fundamental limits

MEV in Binance Builder: Two and Three-Swap Paths Dominate

Empirical data showing that complex multi-hop arbitrage paths aren’t used in production because latency exposure and slippage kill profitability. The real MEV game is optimization on constrained, obvious paths - not finding exotic routes.

Domains: MEV, HFT, market microstructure, empirical analysis

NVIDIA’s approach to MoE training uses network-aware scheduling: route latency-sensitive ops through NVLink (~900GB/s, sub-microsecond) and bandwidth-heavy ops through RDMA. Same principle as HFT network topology optimization, different domain.

Domains: distributed systems, AI infrastructure, network topology, latency optimization

The Pulse of 500+ GPUs: Network Metrics Predict Training Failures

At scale, GPU cluster failures are predicted by interconnect health, not GPU utilization. The B200 generation is so computationally powerful that network becomes the bottleneck - just like what happened in HFT when trading logic got faster than network fabric.

Domains: distributed systems, AI infrastructure, failure prediction, network monitoring

2026-02-21

Arbitrum Timeboost: Protocol-Level MEV Capture Changes the Game

Arbitrum has implemented an auction mechanism for transaction ordering priority, fundamentally shifting MEV extraction from ‘fastest executor wins’ to ‘highest bidder wins.’ This is a critical development for anyone running MEV strategies on L2s - it changes the economics from infrastructure investment to capital allocation and auction game theory.

Domains: MEV, market microstructure, mechanism design, crypto

Why Simple Arbitrage Paths Win: Empirical Evidence from Binance

Research on Binance Builder shows that 2-3 swap arbitrage paths dominate because complexity accumulates execution costs faster than theoretical profit. A valuable lesson in theory versus practice: optimize for fast execution of simple cycles rather than sophisticated pathfinding of complex opportunities.

Domains: MEV, HFT, market microstructure, crypto

Finite-Blocklength Information Theory: Shannon Limits on Low-Latency Communication

Achieving low latency with finite blocklengths requires operating well below channel capacity - you fundamentally trade throughput for speed. This has profound implications for understanding latency arbitrage in financial markets as an information-theoretic problem with theoretical bounds.

Domains: information theory, latency, market microstructure

Microsecond-Scale Queue Priority: When Does Speed Competition Become Wasteful?

Contemporary HFT systems achieve latency in tens of microseconds. At this temporal resolution, queue priority becomes purely technological infrastructure rather than information or skill. Raises fundamental questions about whether continuous markets should shift to discrete-time batch auctions to eliminate socially wasteful speed competition.

Domains: HFT, market microstructure, latency, market design

Monitoring 500+ B200 GPUs: Production Lessons from the Blackwell Frontier

Backend.ai shares operational experience running a 504-GPU B200 cluster, emphasizing that predicting failures requires holistic monitoring beyond just GPU metrics. First real production insights into Blackwell architecture at meaningful scale - valuable for anyone building large-scale distributed training infrastructure.

Domains: distributed systems, GPU, infrastructure, monitoring

2026-02-21

Arbitrum Timeboost: Protocols Capturing MEV Through Priority Auctions

Fundamental shift in MEV extraction economics - Arbitrum now auctions off transaction ordering rights, changing the game from speed competition to auction bidding. Template for other L2s and possibly L1s. Changes MEV from latency optimization to auction game theory.

Domains: MEV, blockchain, market design, auction theory

Finite-Blocklength Information Theory: The Math Behind Latency-Constrained Communication

Shannon’s framework assuming infinite blocklength doesn’t apply to latency-sensitive systems. This paper shows achievable rates fall below classical capacity when delay is constrained - suggests fundamental limits to HFT signal transmission speed.

Domains: information theory, HFT, latency, theoretical limits

Innovation Coding: Rethinking Information Flow Under Delay Constraints

Shifts focus from symbol-level fidelity to information-level utility. In markets, participants don’t need every price tick - they need the ‘innovation’ (unexpected movements). Reconciles why markets can be both efficient and exhibit persistent latency arbitrage.

Domains: information theory, market efficiency, signal processing

MEV in Binance Builder: Why 2-3 Hop Arbitrage Paths Dominate

Empirical analysis of BNB Chain MEV shows arbitrage strongly favors short paths to minimize slippage and latency exposure. Longer paths accumulate execution risk. Complexity kills in MEV extraction - key insight for arbitrage algorithm design.

Domains: MEV, arbitrage, DeFi, algorithmic trading

Operating 500+ GPU Clusters: Lessons from 63-Node B200 Infrastructure

Real operational data showing GPU monitoring alone is insufficient - system-level monitoring (network, power, cooling) is critical. Single failures cascade to entire training jobs. Parallels to HFT infrastructure monitoring requirements.

Domains: infrastructure, ML training, distributed systems, reliability

2026-02-22

MEV in Binance Builder: Two-Swap Dominance

Fresh arXiv preprint showing arbitrage profit concentrates in simple 2-3 swap paths, not complex cycles—latency kills complexity. Direct evidence that speed matters most for simple strategies, which has huge implications for MEV infrastructure design.

Domains: MEV, cryptocurrency, market microstructure, arbitrage

Arbitrum Timeboost: Express Lane Auctions for Transaction Ordering

Arbitrum’s new mechanism for protocol-level MEV capture via priority auctions. Changes the game from pure speed competition to speed + auction strategy. Shows how L2s are evolving beyond first-come-first-served ordering.

Domains: MEV, L2 scaling, mechanism design, Arbitrum

Innovation Coding: Transmitting Less When You’re Faster

Challenges Shannon’s traditional symbol-sequence model for time-sensitive communications. Key insight: lower latency means less innovation accumulates before transmission, so you need to send less data. Potentially foundational for understanding latency-reliability tradeoffs in correlated time series.

Domains: information theory, latency, signal processing

NVIDIA’s Hybrid Expert Parallel for MoE Training

Deep dive on using NVLink (1-2μs) for intra-node + RDMA (5-10μs) for inter-node expert routing in MoE models. The 5-10μs RDMA latency is the bottleneck—exactly where communication speedups would unlock larger expert counts.

Domains: distributed training, MoE, GPU interconnect, RDMA

Monitoring 500+ GPUs: The Pulse of Large Clusters

Production experience running 504 B200 GPUs shows monitoring requires correlating signals across GPU/network/storage/power subsystems. Validates that telemetry systems naturally tolerate some data loss/errors while faster collection enables better failure prediction.

Domains: GPU clusters, monitoring, infrastructure, reliability