OpenAI’s $750 Billion Compute Buildout Now Spans Stargate, Nvidia, 8 GW of Ohio Data Centers and Its Own AI Chip

OpenAI’s reported roughly $750 billion compute-and-cloud spending trajectory through 2030 now sits inside a much larger industrial system. The company is coordinating leased data centers, Stargate partnerships, cloud providers, Nvidia-backed capacity, energy development, multiple accelerator vendors and its first custom inference chip. Chris Malone’s departure is the news peg. The coordination problem is the story.

Why did Chris Malone leave OpenAI?

The verified record establishes that Chris Malone left OpenAI in August 2026 and that the company had reorganized its infrastructure organization earlier in the year. OpenAI confirmed the departure and said the reorganization was intended to support the scale and pace of its work. The record does not establish why Malone left, or whether the departure was voluntary or involuntary.

Malone joined OpenAI in March 2025 after nearly five years at Meta and more than a decade at Google. He initially reported directly to OpenAI President Greg Brockman. Reporting reviewed by POPR Newsroom says his reporting line later moved beneath Sachin Katti after Katti assumed broader leadership responsibilities across infrastructure.

TechCrunch reported that the reorganized structure included Uday Ruddarraju leading the data-center team, Brent Mayo leading data-center build and delivery, and Spas Lazarov leading data-center engineering. The Information separately reported that Malone and Adrian Caulfield co-led a technical engineering and design function focused on AI server-cluster design, cluster scale, facility location and the effects of chip selection on data-center requirements.

Those are reported organizational functions, not all independently published OpenAI titles. The distinction matters because an AI data center is not simply a building containing servers. Accelerator decisions can affect power, cooling, networking, rack design and cluster architecture. Facility location interacts with energy availability and construction timelines. Capacity must arrive when the models and products requiring it are ready to consume it.

The reporting map also has a clear limit. The record supports saying that Malone’s reporting moved from Brockman to Katti and that specialized data-center functions emerged inside the broader infrastructure organization. It does not support saying that Katti replaced Brockman or that every reported leader retains a current formal title from OpenAI. That precision is necessary because the infrastructure story depends on roles and responsibilities that have not all been independently published by the company.

What changed in OpenAI’s infrastructure leadership?

The verified change was a move toward more specialized infrastructure responsibilities under a broader compute strategy. Sachin Katti publicly identifies his role as VP, Compute Strategy & GPT-Infra, with responsibilities spanning compute-capacity expansion, faster deployment, infrastructure economics, return on infrastructure investment, semiconductors, systems and clouds.

Reporting during the reorganization placed multiple infrastructure functions under Katti and placed Katti beneath Brockman. The defensible description is therefore narrower than saying one executive replaced another: Malone’s direct reporting moved from Brockman to Katti while Brockman remained above the infrastructure organization and direct operational responsibilities became more specialized.

This structure becomes more understandable when viewed against what OpenAI is trying to build. The company has to coordinate decisions about capacity, cost, chips, clouds, power, partners and physical delivery at the same time.

That coordination task has a financial as well as an engineering dimension. A data-center organization has to weigh when capacity comes online, how long a lease lasts, what hardware a facility is designed to host, how much power can be delivered and whether a compute commitment can produce useful capability at an acceptable economic return. Katti’s public emphasis on infrastructure economics and return on investment makes that scope relevant without proving any unreported motive for the reorganization.

Is OpenAI’s compute strategy bigger than Stargate?

Yes. OpenAI’s own August 25 strategy statement describes compute as an integrated system spanning data centers and chips, frontier models, its developer platform, consumer products, enterprise products and AI-native devices. OpenAI also says it manages a provider portfolio for capability and economics.

That portfolio names Microsoft, Nvidia, Amazon Web Services, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy and SoftBank. Celestica is separately involved in OpenAI’s custom-silicon program. These companies occupy different positions across cloud computing, accelerators, networking, custom silicon, data-center development, systems integration and energy.

The evidence supports a multi-provider and multi-layered strategy. OpenAI is not merely buying GPUs. It is not merely leasing data centers. And the evidence does not show that it has abandoned Stargate in favor of leasing.

OpenAI is using several procurement and deployment models at once:

• cloud capacity;
• long-duration leases;
• partner-built facilities;
• energy partnerships;
• multiple accelerator suppliers; and
• custom chips.

The result is a distributed industrial network that has to function like a coherent computing platform.

The individual relationships are easier to understand when separated by function. Microsoft, AWS, Oracle, CoreWeave and Cerebras are part of the company’s compute and cloud portfolio. Nvidia and AMD participate at the accelerator layer. Broadcom and Celestica work on custom silicon and systems. SB Energy develops data-center and energy infrastructure. SoftBank is connected through Stargate and its SB Energy investment. The central management problem is not simply adding more partners. It is making capacity, systems, economics and delivery timing work across all of them.

Which OpenAI partners serve which infrastructure role?

OpenAI’s named provider portfolio does not describe one interchangeable group of suppliers. It describes a chain of different capabilities that must work together. Cloud providers and specialized compute companies can make capacity available. Accelerator suppliers shape the hardware available inside that capacity. Data-center developers and energy partners determine where and when facilities can operate. Custom-silicon partners affect the systems OpenAI can deploy in future generations.

The table is a map of reported and first-party-supported roles, not a claim that every company performs the same function or that every relationship has the same commercial terms. It also does not establish that OpenAI controls every layer directly. The company’s own public description is a provider portfolio managed for capability and economics, with direct work where tighter co-design provides leverage.

That distinction helps explain why OpenAI’s infrastructure leadership is a central part of the story. A decision about a future inference chip can affect system design. A decision about capacity can affect lease economics and delivery timing. A decision about a location can affect power availability and the kind of hardware a campus can support. The same organization has to reconcile those choices without treating the cloud, chip and data-center layers as isolated projects.

What does the $750 billion OpenAI figure actually cover?

The approximately $750 billion figure is a reported projection for compute and cloud spending through 2030, not a $750 billion commitment to data-center construction. Wall Street Journal reporting places the projected total at roughly $750 billion, up from an earlier projection of approximately $600 billion.

That evidence boundary is important. The figure covers a broader category than concrete, land and electrical substations. OpenAI’s compute requirements include capacity acquired through outside clouds, accelerator relationships, leased infrastructure, partner-developed facilities and increasingly its own silicon.

Managing the economics of those different forms of compute is therefore part of the infrastructure problem itself. A reported spending projection is not the same thing as a signed construction budget, and the article does not treat it as one.

The figure is still commercially meaningful because it frames the scale of decisions being distributed across the provider portfolio. It does not tell readers how much OpenAI will spend on a particular Ohio building, a particular Stargate facility or a particular chip generation. Those categories should not be merged into one unsupported construction number.

How much Ohio data-center capacity is OpenAI planning?

OpenAI announced on August 17 that it had entered an agreement to secure approximately 8 GW-IT at the PORTS-Pike Technology Campus in Pike County, Ohio, over time. The project is planned capacity delivered in stages, not 8 GW already operating.

SB Energy will build, own and operate the data center, while OpenAI will be the customer under a 20-year lease. OpenAI says the first 800 MW is expected to become available in 2028. The full development depends on infrastructure, permitting, environmental review and financing.

OpenAI projects 35,000 construction jobs during the six-year buildout through 2032 and 2,500 long-term operating jobs. Those are project projections, not present employment totals.

PORTS-Pike is planned around 10 GW of new power capacity, including approximately 9.2 GW of natural-gas generation. That is a power-capacity description, not a claim that 10 GW of new generation is already operating.

The Ohio arrangement illustrates the difference between a signed long-term capacity plan and an operating facility. The first 800 MW is expected in 2028, while the broader 8-GW-IT target is delivered over time. The project depends on physical infrastructure, permits, environmental review and financing. A large projected job count and a large planned power number describe the intended buildout, not current on-site activity.

What must happen before Ohio’s full capacity is operating?

The verified agreement gives OpenAI a path to secure approximately 8 GW-IT over time. It does not make the full capacity operational today. SB Energy is responsible for building, owning and operating the data center, while OpenAI is the long-term customer under a 20-year lease. The first 800 MW is expected to become available in 2028, making the distinction between contracted capacity and active capacity central to accurate coverage.

The project also depends on enabling work outside the data-center buildings themselves. OpenAI identifies infrastructure, permitting, environmental review and financing as conditions for full development. PORTS-Pike’s planned 10 GW of new power capacity, including approximately 9.2 GW of natural-gas generation, describes the scale of energy infrastructure associated with the campus. It does not mean the generating capacity is already energized or that every megawatt is committed to a live OpenAI workload.

The employment figures belong to the same forward-looking category. OpenAI projects 35,000 construction jobs during the six-year buildout through 2032 and 2,500 long-term operating jobs. Those are meaningful project projections, but they are not current staffing numbers. Treating planned capacity, planned power and projected jobs as present conditions would make the Ohio development sound further along than the verified record allows.

What is Nvidia’s role in the Ohio project?

Nvidia is connected to the Ohio project in three distinct roles: AI compute provider, investor in SB Energy and provider of lease-related residual-value credit support.

The Ohio site is expected to exclusively host Nvidia AI compute infrastructure. Nvidia is also investing $1.5 billion in SB Energy. It has provided residual-value guarantees connected to approximately 4.25 GW of initial leased IT capacity and may elect to provide credit support for roughly another 3.8 GW. Reuters reported that the potential support could reach approximately $105 billion.

The approximately 8-GW OpenAI capacity target and the Nvidia guarantee structure are related but not identical. The disclosed guarantee language concerns the initial approximately 4.25 GW of leased IT load, with additional support for roughly 3.8 GW remaining optional. POPR Newsroom does not describe the arrangement as Nvidia guaranteeing the entire construction cost of the Ohio project.

A source-provenance correction also matters here. Material associated with Nvidia’s disclosure was previously hosted by a third party and described in the research ledger as an Nvidia SEC filing. A matching SEC.gov first-party record was not independently located. POPR Newsroom therefore does not represent that third-party document as a first-party SEC source. The guarantee findings remain supported by independently verified reporting.

Keeping the three Nvidia roles separate prevents a second kind of overstatement. Nvidia is not simply a chip vendor at the Ohio site. It is also an investor in the project developer and a provider of specified lease-related credit support. None of those roles changes the published boundary: the disclosed support is tied to phased leased IT capacity, not a blanket guarantee of the entire project.

Why does Nvidia’s credit support need its own boundary?

The Ohio project combines a large capacity target with a more specific financial instrument. Nvidia’s disclosed residual-value guarantees relate to leases for approximately 4.25 GW of initial IT load. The company may, at its sole discretion, provide credit support for approximately another 3.8 GW. Reuters reported that potential support could reach approximately $105 billion. Each number describes part of the arrangement, but none of them is a substitute for the others.

The first boundary is capacity. OpenAI’s approximately 8-GW-IT target is the planned capacity it expects to secure over time at PORTS-Pike. Nvidia’s disclosed guarantee obligations are tied to an initial approximately 4.25 GW, with a separate optional amount. The second boundary is financial. Residual-value support for specified leased capacity is not the same thing as a guarantee of all construction spending or of the full commercial outcome of the campus.

The third boundary is sourcing. The research record notes that a third-party-hosted document was previously described as a first-party Nvidia SEC filing, but a matching SEC.gov record was not independently located. POPR Newsroom therefore does not present that hosted document as a first-party SEC source. The underlying guarantee findings remain supported through the separately verified reporting cited in this report, and the correction protects the distinction between a reported financial structure and an inflated claim about a project guarantee.

Is Stargate being replaced by leased data centers?

No. Leased and partner-built infrastructure can be part of Stargate. In January 2026, OpenAI and SoftBank announced a partnership with SB Energy explicitly as part of Stargate. OpenAI and SoftBank Group each invested $500 million in SB Energy, OpenAI signed a 1.2-GW data-center lease, and SB Energy was selected to build and operate the previously announced 1.2-GW site in Milam County, Texas.

OpenAI described that arrangement as combining its first-party data-center design with SB Energy’s data-center and energy-development capabilities. Stargate can therefore include infrastructure that another company builds and operates while OpenAI leases the resulting capacity.

The Ohio structure pushes the model further. SB Energy again serves as developer, owner and operator. OpenAI becomes the long-term compute customer. Nvidia supplies the AI compute infrastructure while also investing in SB Energy and participating in lease-related credit support.

The evidence does not support a clean distinction between OpenAI-built and outsourced infrastructure. It supports an interconnected model in which design, capital, construction, energy, chips, leases and compute demand are distributed among multiple companies.

That model is important because an announcement of new leased capacity is not evidence that Stargate has been displaced. The verified record supports the opposite, narrower conclusion: Stargate itself can include partner-built and leased infrastructure. OpenAI appears to be using several acquisition and deployment models at once rather than choosing one exclusive path.

Why is OpenAI building its own AI chip?

OpenAI and Broadcom unveiled Jalapeño on June 24, 2026. OpenAI describes it as its first Intelligence Processor and first custom inference chip, designed around the inference requirements of its large language models.

OpenAI is responsible for the custom accelerator design. Broadcom provides silicon implementation, networking and connectivity technologies. Celestica contributes board, rack and system expertise. Initial deployment is planned by the end of 2026, followed by expansion across future generations.

Richard Ho, identified in current reporting as OpenAI’s vice president of hardware, is associated with the custom-chip program. Jalapeño means OpenAI is not only diversifying which external accelerator it buys. It is adding a vertically integrated silicon program to a portfolio that already includes Nvidia, AMD and other compute providers.

A custom processor affects more than semiconductor procurement. OpenAI’s infrastructure organization has been dealing with the relationship between chip choices and data-center requirements. The chip and the building cannot be managed as completely separate systems.

Jalapeño therefore changes the infrastructure story in a specific way. OpenAI is adding an internally designed accelerator to a portfolio that still includes external vendors and cloud providers. The record supports that portfolio expansion. It does not support predicting how quickly the custom chip will displace any named outside supplier or how much it will change OpenAI’s future compute cost.

How does Jalapeño connect chips to data-center decisions?

OpenAI describes Jalapeño as its first Intelligence Processor and first custom inference chip, designed around the inference needs of its large language models. OpenAI is responsible for the accelerator design. Broadcom provides silicon implementation, networking and connectivity technologies. Celestica contributes board, rack and system expertise. The initial deployment target is the end of 2026, followed by expansion across future generations.

Those roles show why custom silicon is an infrastructure question rather than a narrow semiconductor announcement. An accelerator’s design affects the system around it, including networking, racks, power, cooling and how it can be deployed with data-center partners. The sealed record identifies the relationship between chip choices and data-center requirements as part of the technical engineering and design work considered during the reorganization.

Jalapeño gives OpenAI another way to co-design compute for its models, but it does not erase the rest of the provider portfolio. Nvidia, AMD, cloud providers, data-center developers and systems partners remain part of the verified picture. The defensible conclusion is that OpenAI is broadening its options across the stack. The record does not establish a date on which its custom chip will replace any external accelerator supplier.

Does OpenAI need a new management layer for compute?

The verified facts support that as POPR Newsroom’s analysis, not as terminology OpenAI has adopted. OpenAI increasingly needs executive coordination across capacity, economics, chips, clouds, power, partner relationships and physical data-center delivery.

Microsoft, AWS, Oracle, CoreWeave and Cerebras can supply different forms of compute. Nvidia and AMD participate at the accelerator layer. Broadcom and Celestica participate in custom silicon and systems. SB Energy participates in data-center and energy development. SoftBank is connected through Stargate and SB Energy. Nvidia appears simultaneously as accelerator supplier, SB Energy investor and provider of lease-related credit support around the Ohio project.

Each relationship can make sense independently. The infrastructure organization has to make them make sense together. That is a different management problem from simply securing as many GPUs as possible.

OpenAI must consider when capacity becomes available, what it costs, which chips it uses, where it operates, how quickly it can be energized, how systems interconnect and how much useful intelligence the company can extract from the capital deployed. Katti’s public emphasis on infrastructure economics and return on investment is notable in that context.

Compute scarcity may remain a technological constraint. Compute coordination is increasingly an organizational one.

This is the analytical center of the report. It is a POPR Newsroom interpretation of the verified facts, not an OpenAI label for its organization. The evidence supports saying that coordination is becoming a central management challenge. It does not support presenting an internal analytical phrase as OpenAI’s own strategy or claiming that a leadership change has solved or caused the challenge.

Does executive turnover prove that OpenAI has an infrastructure crisis?

No. The verified record establishes executive turnover, but it does not establish a common cause or an infrastructure crisis.

TechCrunch reported that Malone’s departure joined more than a dozen executive departures during 2026, citing a Business Insider count of 13. Other verified transitions include Denise Dresser’s planned departure from the chief revenue officer role, OpenAI’s August 13 appointment of Dali Rajic as her successor, Brad Lightcap’s departure after a move into a special-projects role, and Fidji Simo’s move out of full-time executive leadership for publicly disclosed health reasons while continuing in an advisory capacity.

Those events establish a turnover pattern. They do not establish that Malone and the other executives left because of the same organizational problem, that the departures demonstrate an infrastructure crisis or that they collectively prove instability inside OpenAI.

Malone’s departure intersects with an infrastructure reorganization, but temporal proximity can invite a causal story that the evidence does not provide. OpenAI has not disclosed why Malone left, and POPR Newsroom does not assign one.

That distinction is especially important in a report about physical infrastructure. A reorganization, a departure and a large capital plan can appear together without supplying evidence for a single causal narrative. The verified story is substantial without that inference: OpenAI is expanding a complex compute system while reorganizing the people responsible for parts of it.

Does OpenAI’s potential IPO explain the reorganization?

No. The potential IPO is relevant context for the economics of OpenAI’s compute system, but the evidence does not establish that IPO preparation caused the infrastructure reorganization.

OpenAI disclosed on June 8 that it had confidentially submitted a draft Form S-1 registration statement to the U.S. Securities and Exchange Commission for a possible initial public offering. The company also said it had not decided when to proceed. OpenAI has not established a fixed 2027 IPO date.

A company contemplating public ownership while projecting extraordinary compute requirements will face scrutiny of capital requirements, infrastructure economics and returns. That makes the filing relevant to the commercial context. It does not put IPO preparation in the causal chain behind Malone’s departure or the reorganization without evidence that has not been established.

The same discipline applies to private-company valuation figures sometimes attached to OpenAI coverage. The sealed record excludes a reported approximately $852 billion private valuation from the core infrastructure story because it is neither an announced IPO valuation nor necessary to explain the verified compute strategy.

What does the OpenAI infrastructure record not establish?

The report does not establish that Chris Malone was fired, why he left, or that his departure resulted from the earlier reorganization. OpenAI confirmed that he left and that it had reorganized the infrastructure organization, but the causal link between those events has not been published. A personnel departure can be the news peg without becoming evidence for an unverified motive.

The report also does not establish that Stargate is failing, that OpenAI has abandoned Stargate for leasing, or that the company is in an infrastructure crisis. The Texas SB Energy arrangement was explicitly described by OpenAI as part of Stargate even though SB Energy builds and operates the facility while OpenAI leases capacity. That fact makes a simple build-versus-lease narrative inaccurate.

Finally, the report does not establish that all executive transitions in 2026 share one cause, that a potential IPO caused the reorganization, that all 8 GW of Ohio capacity is built or identically guaranteed, or that $750 billion is a data-center construction budget. Those exclusions are not minor caveats. They are the boundaries that keep a large, complex infrastructure story from becoming a misleading one.

What is the larger OpenAI infrastructure story?

The larger story is the physical and financial system underneath OpenAI’s software products. Users encounter ChatGPT, developers encounter APIs and models, and enterprises encounter products. Underneath those interfaces, OpenAI is assembling a system whose components increasingly resemble an industrial supply chain.

There are leased data centers in Texas and Ohio, cloud providers and specialized compute partners, Nvidia accelerators and an OpenAI-designed processor being implemented with Broadcom and Celestica. There are power-generation requirements measured in gigawatts, energy developers, long-term leases, credit-support structures and enormous projected compute expenditures.

There is also an organization responsible for making those pieces work together. That is why Chris Malone’s departure matters without needing to become a story about crisis.

OpenAI reorganized its infrastructure leadership before Malone’s exit. Malone’s reporting moved from Brockman to Katti. Specialized data-center functions emerged within the broader infrastructure structure. Katti’s remit explicitly includes capacity, deployment speed, economics, semiconductors, systems and clouds.

Meanwhile, OpenAI’s physical compute footprint keeps expanding. The company is not merely building bigger models anymore. It is building the industrial system required to keep building them.

With reported compute and cloud spending projected at roughly $750 billion through 2030, the central infrastructure question is becoming as much managerial and economic as technical: can OpenAI turn an enormous network of clouds, chips, leases, power projects, custom silicon and infrastructure partners into one economically coherent compute machine?

Malone’s departure is the news. The system OpenAI now has to manage is the story.

Fact Summary

Chris Malone left OpenAI in August 2026 after joining the company in March 2025. OpenAI confirmed his departure and said it had reorganized its infrastructure organization earlier in 2026 to support the scale and pace of its work. Reporting shows Malone’s direct reporting line moved from Greg Brockman to Sachin Katti. The evidence does not establish why Malone left or whether the reorganization caused his departure.

OpenAI’s infrastructure strategy now spans multiple cloud and compute providers, long-term data-center leases, partner-built infrastructure, energy relationships, multiple accelerator vendors and custom inference silicon. OpenAI itself describes data centers, chips, frontier models, developer infrastructure, products and devices as an integrated full-stack system.

OpenAI has a 1.2-GW lease with SB Energy in Texas that it explicitly identifies as part of Stargate. In Ohio, OpenAI has agreed to secure approximately 8 GW-IT of capacity over time at PORTS-Pike through a 20-year lease with SB Energy. The Ohio project is planned around 10 GW of new power capacity, including approximately 9.2 GW of natural-gas generation.

Nvidia has residual-value guarantees connected to approximately 4.25 GW of initial leased IT capacity and may elect to support roughly another 3.8 GW. Reuters reported that potential support could reach approximately $105 billion. Nvidia is also investing $1.5 billion in SB Energy. POPR Newsroom does not characterize the third-party-hosted document listed in the underlying source ledger as a first-party SEC.gov filing.

OpenAI and Broadcom unveiled Jalapeño in June 2026. OpenAI describes it as its first custom inference chip, with Broadcom providing silicon implementation, networking and connectivity technologies and Celestica contributing board, rack and system expertise. Initial deployment is planned for the end of 2026.

Wall Street Journal reporting places OpenAI’s projected compute and cloud spending through 2030 at approximately $750 billion. That is a broader compute-and-cloud projection, not a $750 billion commitment to data-center construction.

OpenAI has also confidentially submitted a draft Form S-1 for a possible IPO, but the company has not established when it will proceed. The evidence does not establish that IPO preparation caused the infrastructure reorganization.