OpenAI's 1B-User Habitat Reveal Confirms a Second AI Capex Wave—And the Better Storage Trade Isn't Seagate

OpenAI's Habitat blog is short on product drama and long on architecture. The company reveals a storage tier running 70 million requests per second, serving more than 500 petabytes across almost 40 geographic regions for over a billion weekly ChatGPT users. The Rust rewrite that handled 95% of production traffic by Q2 2026 made the service 6× more CPU-efficient and 15× more memory-efficient than the Python original—an order-of-magnitude density win that is, in dollar terms, a storage-and-compute capex offset.

The architecture is the news. Habitat is a NoSQL object-and-edge model running on Azure Cosmos DB, with change-data-capture feeding Databricks, Rockset and Kafka for the offline secondary view. In plain terms: OpenAI has built its primary data layer as software-defined, object-storage-shaped infrastructure with multi-region replication, blob storage underneath, and Valkey caches in front. That stack is where the storage-tier capex is being spent, and it is exactly the workload pattern Micron's 245TB 6600 ION was built for.

ChatGPT crossed 1 billion monthly active users in June 2026, the fastest any app has reached that line, per Reuters' reading of the data. Three months later, OpenAI published what is effectively the storage-tier bill for that milestone: 500 PB online, almost 40 regions, 70M requests/sec. Separately, The Information reports OpenAI is shopping as much as 5 exabytes of data-center storage as part of the Stargate buildout.

The exabyte-scale procurement is the load-bearing number. At industry blended pricing of roughly $20/TB for capacity-class HDD and $80–120/TB for QLC SSD, a 5 EB build implies a $100B+ multi-year storage spend—non-trivial against NVIDIA's $30B OpenAI equity commitment and the $1.5B equity NVIDIA put into SoftBank's SB Energy on Aug 17 to back the Ohio site.

Seagate is not the wrong bet; it is the second-best one. Q3 FY26 gross margin reached 47.0%, up from 36.2% a year earlier; 88% of exabyte shipments and 80% of revenue are now data-center. Management has nearline capacity almost fully allocated through calendar 2027 on build-to-order contracts, and Mozaic 4 lifts per-drive capacity to 44 TB. Western Digital trades at 17× trailing earnings against Seagate's 62×—cheaper on nearline volume leadership plus a longer-dated HBDT roadmap that targets dual-pivot and high-bandwidth drives.

The structural problem for the HDD-pure trade is power and floor space. Micron's 245 TB 6600 ION, shipping since May 5, 2026, cuts energy use 84× and rack count 82% versus an HDD deployment of equivalent raw capacity, and delivers 435× better throughput per watt on object-storage workloads. In a 1 EB deployment, an HDD build consumes 1.9× more energy than the 6600 ION. AI data-center operators are hitting gigawatt-scale power constraints; the storage tier that solves the constraint wins the share.

The same day OpenAI published the Habitat post, Fidji Simo—OpenAI's former CEO of AGI Deployment—joined the board of Nscale, the UK-based \"vertically integrated AI infrastructure\" operator running liquid-cooled data centers across Norway, the UK and a planned Harjavalta, Finland site. Nscale's pitch is software-plus-compute-plus-power under one roof, and its anchor customer is OpenAI through Stargate Norway. NVIDIA is the chip provider; SoftBank is the capital partner.

That stack is what investors should underwrite. NVIDIA's Aug 17 $1.5B equity in SB Energy—paired with the Feb 27 $30B OpenAI investment at a $730B pre-money valuation—ties compute, power real estate and storage procurement into one equity chain. OpenAI's $50B from Amazon adds a third hyperscaler anchor. The structural direction of travel is that the storage tier follows the equity stack, not the unit-shipment cycle.

Three near-term catalysts will test the thesis in the next two quarters. First, Seagate's Q1 FY27 print—guided to $4.1B revenue at the midpoint on July 28, against consensus of $3.75B—will show whether the 2027 sold-out book survives a quarter of exabyte procurement. Second, NVIDIA's Q3 FY27 earnings in mid-November will disclose any incremental OpenAI-related commitments beyond the $30B/$1.5B already on the tape. Third, the Stargate Norway site's first 100,000 GPUs (target end-2026) and the Ohio SB Energy buildout will quantify how much storage capex follows each gigawatt of compute.

The 2028 inflection is harder to ignore. AI agent workloads—OpenAI's managed Agents API launched earlier this month—turn chat sessions into persistent per-user state, multiplying object-store writes per active user. If agent adoption follows ChatGPT's curve toward 1B users, online storage demand per user rises materially; if it plateaus, Seagate's 2028 allocation overhang returns. The 12- to 36-month read is that QLC SSDs keep eating the warm tier while HAMR HDDs retain the cold, and the NAND/DRAM makers carry the higher gross margin per byte.