Keep Tier 0 on‑prem

OpenAI’s $16 billion Michigan data-center fight, Oracle’s debt-backed AI expansion and Cushman & Wakefield’s warning that India’s AI data-center boom is constrained by power and land all landed within days of each other this week. Together, those reports did not change the physics of low-latency trading, but they did change the backdrop around infrastructure decisions. The question for banks and market firms is no longer whether cloud has uses. It is which workloads can tolerate shared infrastructure and which cannot. The answer many infrastructure teams already use is a tiered one. Exchange-adjacent systems stay on on-premises or colocation infrastructure. Telemetry, replay and some control-plane functions can be split across hybrid estates. Training, simulation and other elastic jobs can move to cloud or sovereign AI farms. This week’s reporting added fresh evidence for keeping that line hard. ### Why did a Michigan zoning fight matter to trading architects? Saline Township, Michigan, became part of the infrastructure debate when OpenAI, through its Stargate effort with Oracle, pushed ahead with a $16 billion data-center project after a zoning dispute, according to Attack of the Fanboy. The report said the township had initially denied zoning changes and that Related Digital then sued, leading to a settlement that cleared the way for construction. That episode mattered because it showed how politically contested premium AI capacity has become. A large AI site can now involve local land-use fights, utility questions and legal pressure before a single workload is placed. For firms deciding where to run latency-critical systems, that is a reminder that AI-scale infrastructure is being built for capacity and throughput, not for deterministic exchange proximity. (attackofthefanboy.com) ### What did Oracle’s financing say about the cloud side of the equation? Fast Company reported on May 23 that Oracle’s AI expansion is being financed heavily through private credit, describing debt as both fuel for the boom and a source of risk. The article focused on how capital-intensive AI infrastructure has become as providers race to add capacity. That matters for workload placement because debt-heavy expansion tends to favor large, utilization-hungry facilities. (attackofthefanboy.com) Those facilities are well suited to model training, batch analytics and developer workloads that can absorb queueing, distance and variable scheduling. They are not designed around the bounded jitter and venue adjacency that low-latency order handling requires. That inference follows from the different operating requirements of cloud-scale AI sites and exchange-edge systems. (fastcompany.com) ### Why does India’s data-center buildout point the same way? Communications Today, citing Cushman & Wakefield, reported on May 23 that India’s AI data-center boom is unlikely to create oversupply because power shortages, land constraints and tighter regulation are acting as brakes. Other outlets carrying the same report used similar language about natural limits on overbuilding. (fastcompany.com) Those constraints reinforce a simple point: premium compute is becoming more local, more power-bound and more selective. When capacity is scarce or politically constrained, firms tend to reserve it for workloads that need scale rather than for workloads that need deterministic microsecond behavior. That is one reason many architects separate hot-path trading from AI and development estates even when both sit inside the same broader firm. (communicationstoday.co.in) ### Where do security risks fit into the placement decision? The Register reported on May 23 that recent Linux flaws dubbed Dirty Frag, Copy Fail and Fragnesia pointed to a “worrisome” trend around kernel memory-management weaknesses. Separate reporting from BleepingComputer and others said Fragnesia allowed local users to gain root privileges and prompted patching across Linux distributions. (communicationstoday.co.in) For latency-sensitive systems, that supports keeping Tier 0 small. The fewer agents, dependencies and privileged control-plane components attached to execution-adjacent hosts, the easier those systems are to patch, validate and roll back. The security argument is not that cloud is inherently unsafe; it is that the most latency-critical estate benefits from fewer moving parts and tighter operational control. (theregister.com) ### So what does the three-tier model actually look like? Tier 0 usually means exchange connectivity, feed handling, order gateways and deterministic pre-trade controls on on-premises or colocation infrastructure. Tier 1 usually covers replay, telemetry aggregation, surveillance enrichment and diagnostics in a controlled hybrid setup. Tier 2 usually includes model training, CI pipelines, simulation, documentation and batch analytics in cloud or sovereign AI environments. This week’s reporting did not invent that framework, but it gave fresh reasons to preserve it. (theregister.com) Oracle, OpenAI, Related Digital and Cushman & Wakefield are likely to remain part of that conversation as more AI capacity comes online and more firms revisit placement rules in 2026. The next concrete inputs will come from additional project disclosures, financing updates and capacity reports tied to Oracle’s buildout, the Saline Township site and Cushman & Wakefield’s market tracking. (attackofthefanboy.com)