Thoughts on low latency trading if exchanges went full cloud

Technical barriers to “full cloud” exchanges

• Major obstacle: lack of true hardware multicast in hyperscale clouds. Existing “multicast” is software fan‑out over unicast, adding latency and jitter.
• Ultra‑low‑latency (ULL) setups depend on purpose‑built networks where 10–40Gbps+ multicast can saturate links with predictable behavior and minimal queuing.
• Cloud SDNs hide L2, ARP, and multicast by design, limiting the ability to build these specialized fabrics.
• Cloud FPGAs exist but typically as compute accelerators without direct network access; trading FPGAs need line‑rate NIC control and protocol “abuse.”
• Concerns about cloud reliability vs exchange uptime; cloud outages and noisy neighbors are seen as unacceptable for matching engines.

Fairness, latency, and colocation

• Exchanges use equal‑length fiber and carefully engineered networks to approximate “provable fairness” within a colo; colocation remains the dominant model.
• Moving matching engines into EC2‑style environments would make it harder to prove identical latency, and would invite games around instance placement, ping‑mapping, and potential DDoS on competitors’ hosts.
• Some argue clouds could create special regions or Outposts racks with symmetric connectivity, effectively recreating colo “inside” the cloud.

Role and impact of HFT / ULL trading

• Explanations for newcomers: speed lets firms react first to news, arbitrage across venues, and gain time priority in price‑time order books; profits are tiny per trade but huge in volume.
• Supporters say HFT competition tightens spreads, improves retail execution, and replaces the older, wider‑spread floor‑trader regime.
• Critics see HFT, especially latency arbitrage, as rent‑seeking that extracts value from slower participants and drives an arms race in hardware and proximity.
• Empirical studies (cited in-thread) claim latency arbitrage is frequent and costly, though definitions and interpretations are disputed.

Batch auctions and artificial latency

• Several propose discrete batch auctions (from milliseconds to seconds) or random/jittered delays to blunt latency advantages and front‑running.
• Others note many markets already use opening/closing auctions and volatility halt auctions; experiments with frequent auctions showed mixed impact.
• Concerns: batch systems can be gamed (e.g., outsized orders, pro‑rata matching), may increase spreads, complicate multi‑asset strategies, and shift from speed games to size/strategy games.

Economics and incentives of moving to cloud

• Exchanges earn substantial revenue from colocation, cross‑connects, and proprietary feeds; moving to cloud would hand margin to cloud providers and reduce control.
• Some see little operational or cost upside for mature exchanges that already have specialized staff and hardware.