# AI infrastructure stocks — X 热门讨论 (2026-09-28 16:24 UTC)
## @BrettHarrison (Brett Harrison) · 09-28 13:38 · ♥55 ↻5 💬2 Profitability in HFT depends as much on how a firm manages operational risk that resists automation or AI solutions as on latency optimization and model predictivity. The most mundane examples are actually the hardest:
• New listings. Every back- and front-office system needs to be prepared for a new listing when details are made available a few hours ahead of the first trade. Initial versions of market data infrastructure and trading systems are commonly built on the assumption that all symbols are known at startup. If a symbol’s metadata is unknown until listing day, a firm must support intraday additions without restarting systems en masse, or risk firm-wide market outages. One solution is to ingest all existing and new metadata definitions within a stream processing framework, treating symbology in a similar fashion to market data. • Corporate actions. Dividends, splits, ticker changes, and rights offerings are all examples of equity corporate actions that cause discontinuous changes in pricing and can be catastrophic for P&L if missed. Handling simple corporate actions like dividends can be highly complex, e.g. trading in countries in which stocks trade ex-dividend before the amount is determined. Options corporate actions are more complicated, often producing adjusted series with nonstandard deliverables and new root symbols that trade alongside standard contracts. • Position reconciliation. Traditional exchanges and clearing firms do not offer real-time feeds of customer positions. HFTs need to calculate their own view of positions across all trading venues, which can diverge from sources of truth due to trade corrections, busts, missed packets, metadata errors, contract multiplier changes, and clearing firm give-ups. A single mismatched position can lead a firm to have an improperly hedged portfolio, fail to execute PFOF trades promised to a retail broker channel, or worse. • Market data staleness. A key problem when subscribing to real-time market data is distinguishing between a short period with no market activity and a networking issue that causes packets to be silently dropped. Continuing to provide two-sided liquidity in a symbol where the book or the underlying is stale can rapidly erase profitability. TCP guarantees delivery, and most UDP feeds carry sequence numbers and heartbeats that expose gaps, but neither protocol can indicate whether silence in a particular symbol implies inactivity or a problem. The HFT industry standard is to pre-configure systems with expected update intervals per symbol, issuing warnings or automatic shutoffs when staleness thresholds are breached.
An LLM trained on an HFT’s internal codebase will struggle to build software that manages the above day-to-day events with required safety for production. Large HFTs have established operational competitive advantage via trial and error over years and at large costs, relying on highly specific personnel and manual technical workflows. This class of firm may surprisingly have the strongest disincentives to explore or improve LLMs for operational trading, even internally. https://x.com/BrettHarrison/status/2104566331890610655
## @EhrmantrautCap_ (Ehrmantraut Capital) · 09-27 21:38 · ♥30 ↻7 💬5 We're not quite there yet, but the inflection point towards positive ROIC would not only be bullish for hyperscalers themselves, but imo also AI infrastructure stocks like $AMD, $INTC, $MU, $SNDK, $AVGO, etc., through multiple channels:
Positive ROIC -> higher cashflow resulting from higher AI revenue -> more money to spend on compute
Positive ROIC -> easier to justify spending more on compute towards shareholders -> more likely to issue debt to spend on compute
Positive ROIC -> higher cashflow resulting from higher AI revenue -> better financial health -> easier to issue more debt to spend on compute > 引用 @pequityresearch: Goldman Sachs: AI ROIC
"The hyperscalers' current estimated Al revenues remain below the amount that would allow them to break even on their capex, but those revenues are growing quickly and revenue backlogs are sizable."
0% ROIC --> $308 billion hyperscaler AI revenue 10% ROIC --> $417 billion hyperscaler AI revenue 15% ROIC --> $526 billion hyperscaler AI revenue 20% ROIC --> $636 billion hyperscaler AI revenue
$GOOGL $MSFT $ORCL $META $AMZN https://x.com/EhrmantrautCap_/status/2104324936605065407