OpenAI and Broadcom unveiled Jalapeño in June 2026 as OpenAI’s first custom AI inference processor, a co-designed accelerator intended for large language model workloads across ChatGPT, Codex, the API, and future agent-style products. The announcement, amplified by Cloud Wars and corroborated by OpenAI’s own release, is less about one chip than about a company trying to escape the gravitational pull of commodity infrastructure. OpenAI is no longer behaving like a model lab that rents enough cloud capacity to survive the next launch. It is acting like a platform company that believes the economics of intelligence will be decided in silicon, networking, power, and software together.
According to OpenAI and Broadcom, Jalapeño is the first AI accelerator in a multi-generation compute platform, not a one-off science fair ASIC. OpenAI says it designed the architecture around its understanding of LLM serving, while Broadcom helped turn that design into manufacturable hardware and production systems. Celestica is also part of the industrialization story, helping with board, rack, and integration work.
That makes the chip strategically different from a generic “Nvidia alternative” headline. OpenAI is not simply trying to buy cheaper compute. It is trying to make the hardware reflect the shape of its own models, kernels, serving systems, and product roadmap.
The bet is that inference workloads are now predictable and valuable enough to justify custom silicon. If OpenAI knows where its models are going, how its products behave, and what bottlenecks dominate real-world serving, it can encode those assumptions into hardware. That is risky, because hardware freezes choices that software can revise overnight. But it is also how platform companies turn scale into margin.
OpenAI’s problem is not that Nvidia makes bad chips. It is that Nvidia makes chips everyone else wants too. When the same scarce hardware is being chased by hyperscalers, model labs, sovereign AI projects, startups, research groups, and enterprises, even the best supplier relationship becomes a constraint.
Broadcom’s role is important precisely because it specializes in the less glamorous but deeply consequential world of custom silicon and high-scale networking. The AI race is not only about matrix multiplication. It is about moving data, keeping racks fed, cooling dense systems, and making clusters behave as reliable products rather than heroic engineering projects.
OpenAI’s earlier 10-gigawatt accelerator collaboration with Broadcom, announced in October 2025, already pointed in this direction. Jalapeño gives that partnership a named chip and a clearer narrative. The company wants custom accelerators, custom systems, and enough power-backed infrastructure to serve intelligence at industrial scale.
Every ChatGPT conversation costs something. Every Codex request consumes compute. Every API customer that moves from demo to production converts AI from a capital-intensive bet into an operating-cost challenge. The more successful OpenAI becomes, the more punishing inference economics become unless the company can lower the cost per useful answer.
That is why performance per watt is such a loaded claim. OpenAI says early tests show Jalapeño delivering substantially better performance per watt than current state-of-the-art AI processors. The company has not provided enough public detail for outsiders to verify that claim independently, so it should be treated as vendor positioning until production deployments prove it.
Still, the direction is obvious. In a world where power availability is becoming as important as chip availability, a more efficient inference processor is not merely a nice engineering win. It is a way to stretch data center capacity, reduce serving costs, and make heavier AI products commercially viable.
OpenAI appears to be reaching for a version of that logic. The model architecture, inference runtime, product surface, API behavior, and chip design can all inform one another. If that loop works, OpenAI can optimize for the workloads it actually runs rather than for the average needs of the entire accelerator market.
But OpenAI is not Apple. It does not control the end-user device, the operating system, the browser, the enterprise desktop, or the smartphone distribution layer. Its “device” is the data center, and its customer experience is mediated through apps, APIs, cloud partners, and increasingly through other companies’ software.
That makes the integration play both more abstract and more fragile. Apple could use custom silicon to make a MacBook feel faster and last longer on battery. OpenAI has to use custom silicon to make remote intelligence feel faster, cheaper, and more reliable across millions of unpredictable workloads.
But chip development is full of cliffs hidden behind triumphant tape-out announcements. A working sample is not the same as reliable high-volume deployment. A benchmark does not equal fleet economics. A rack that runs well in a lab can behave differently when thousands are deployed under real customer load.
Broadcom reduces the risk because it knows the ASIC business, the networking layer, and the hyperscale supply chain. But even Broadcom cannot repeal physics, packaging constraints, thermal limits, yield realities, or the brutal calendar of semiconductor manufacturing.
That is why OpenAI’s language about a “multi-generation compute platform” matters. The first chip does not have to be perfect if it establishes the design loop. The real prize is learning how to build Jalapeño’s successors faster, more efficiently, and more tightly aligned with OpenAI’s model roadmap.
That creates an interesting tension. Microsoft has its own AI silicon work, including Azure Maia, and it has every reason to reduce dependency on any single accelerator vendor. OpenAI building custom inference chips could complement Microsoft’s ambitions if it increases available capacity and lowers costs for shared services. It could also complicate the stack if OpenAI’s preferred hardware path diverges from Azure’s broader infrastructure strategy.
The likely near-term outcome is coexistence rather than conflict. Hyperscale AI is becoming too large for a single chip family, a single cloud architecture, or a single procurement channel. Nvidia GPUs, AMD accelerators, cloud-native silicon, and custom ASICs can all coexist because demand is expanding faster than supply can rationalize itself.
Still, Jalapeño signals that OpenAI wants more leverage in the relationship. A model company that depends entirely on someone else’s cloud and someone else’s chips has limited bargaining power. A model company with its own accelerator roadmap becomes a more complicated partner.
The practical relevance is upstream. If Jalapeño works, it could change the cost and latency profile of OpenAI-powered services. That eventually affects API pricing, availability, response times, throughput ceilings, and the kinds of AI features vendors can afford to bundle into productivity software.
For administrators and architects, the key question is not whether Jalapeño beats an Nvidia GPU in an isolated benchmark. The question is whether OpenAI can use custom inference hardware to make high-volume AI services more predictable. Enterprises do not merely need dazzling demos; they need service levels, compliance controls, capacity planning, and prices that survive budget season.
That is where custom silicon could matter most. If OpenAI can lower its marginal inference costs, it has more room to support longer context windows, heavier agent workflows, faster code generation, and richer multimodal experiences without turning every customer interaction into an expensive compute event.
Software companies are used to shipping, measuring, and updating. Semiconductor programs demand earlier commitments and tolerate fewer late pivots. A chip designed around today’s inference assumptions may look brilliant if the model roadmap follows expectations, or awkward if a new architecture changes the bottlenecks.
That is the core strategic gamble. OpenAI’s confidence comes from its privileged view of its own workloads. It knows what users ask, how models respond, where latency accumulates, what kernels dominate, and what future products it wants to serve. Most chip vendors have to generalize across customers. OpenAI can specialize.
Specialization is powerful, but it narrows escape routes. If OpenAI’s future models require very different memory behavior, sparsity patterns, routing mechanisms, or runtime orchestration, the hardware roadmap has to keep up. The silicon game rewards conviction, but it punishes the wrong kind of certainty.
The October 2025 OpenAI-Broadcom announcement spoke in gigawatts, not merely chips. That language is revealing. AI infrastructure is now being planned at a scale where electricity, substations, cooling, land, supply chains, and network fabrics are strategic assets.
This is why the Stargate-style data center narrative matters. Whether every announced project arrives exactly as described is less important than the direction of travel. The leading AI companies are trying to secure multi-year access to the physical substrate of intelligence.
For users, that sounds remote until it isn’t. Capacity shortages become waitlists, throttling, higher prices, slower features, regional limits, or degraded service quality. Abundant inference capacity becomes better products. The chip is only one piece, but it is a piece OpenAI increasingly does not want to leave entirely to others.
OpenAI itself is unlikely to abandon GPUs. Frontier model development still benefits from flexible hardware and enormous existing tooling. Even if Jalapeño becomes a major inference engine for OpenAI’s own products, the company will likely continue using a blend of Nvidia, other accelerators, and partner infrastructure.
The more realistic threat to Nvidia is not a single chip. It is workload segmentation. As inference volumes grow, the most repetitive and economically sensitive workloads become attractive targets for custom silicon. GPUs keep the frontier moving; ASICs harvest the stable high-volume paths.
That is exactly how mature computing markets evolve. General-purpose hardware dominates early because flexibility matters most. As patterns stabilize and scale increases, specialized hardware starts carving out profitable lanes. Jalapeño is OpenAI’s claim that LLM inference has reached that stage.
The same is true for Windows developers integrating AI into desktop apps, enterprise tools, and automation workflows. The chip does not change how a developer calls an API. It could change what those APIs can economically promise.
That distinction is important because AI infrastructure stories often get consumed as spectacle. The real developer question is mundane: can I build a feature on this service and trust that it will be fast, available, and affordable at scale? If custom inference hardware improves that answer, it matters even if developers never touch the silicon directly.
There is also a competitive angle. If OpenAI can lower the cost of serving its models, rivals may have to respond through their own silicon partnerships, cloud optimizations, or pricing changes. Developers benefit when infrastructure competition becomes product competition rather than merely benchmark theater.
A company that controls more of its stack can argue that it has a durable moat. It can point to models, distribution, data center capacity, hardware design, and operational learning as mutually reinforcing assets. That is a more compelling story than “we rent GPUs and serve a chatbot,” especially if public-market ambitions ever become real.
But vertical integration also raises the stakes. OpenAI’s capital needs become larger. Its execution burden becomes heavier. Its exposure to supply chain risk, energy politics, and hardware depreciation becomes harder to ignore.
The company’s public narrative has long been about making advanced AI broadly available. Jalapeño gives that mission an industrial form. Broad access requires abundant compute; abundant compute requires hardware, power, and money at a scale few companies can assemble.
Production inference will expose the hard parts. Can the chip handle real traffic patterns? Can OpenAI’s software stack route workloads intelligently? Can the systems run reliably at scale? Can the company measure enough savings to justify the engineering and capital commitment?
If the answer is yes, Jalapeño becomes more than a symbolic Nvidia hedge. It becomes the first step in a feedback loop where OpenAI’s models shape its chips and its chips shape its models. That would be a meaningful structural advantage.
If the answer is mixed, the company still learns. A first-generation custom accelerator can disappoint in some dimensions and still teach the organization how to build the next one. In silicon, the first product is often less important than the roadmap it enables.
OpenAI’s Chip Is Really a Claim on the Future Cost of Intelligence
Jalapeño is being introduced as an inference chip, not a training monster. That distinction matters because inference is where AI becomes a daily utility rather than a research project. Training produces the model; inference serves every prompt, code completion, image request, tool call, and agentic workflow that users expect to feel instantaneous.According to OpenAI and Broadcom, Jalapeño is the first AI accelerator in a multi-generation compute platform, not a one-off science fair ASIC. OpenAI says it designed the architecture around its understanding of LLM serving, while Broadcom helped turn that design into manufacturable hardware and production systems. Celestica is also part of the industrialization story, helping with board, rack, and integration work.
That makes the chip strategically different from a generic “Nvidia alternative” headline. OpenAI is not simply trying to buy cheaper compute. It is trying to make the hardware reflect the shape of its own models, kernels, serving systems, and product roadmap.
The bet is that inference workloads are now predictable and valuable enough to justify custom silicon. If OpenAI knows where its models are going, how its products behave, and what bottlenecks dominate real-world serving, it can encode those assumptions into hardware. That is risky, because hardware freezes choices that software can revise overnight. But it is also how platform companies turn scale into margin.
The Nvidia Dependency Was Always a Symptom, Not the Disease
Much of the coverage naturally frames Jalapeño as a shot at Nvidia. That is true as far as it goes, but it misses the larger economic disease: every major AI company is trapped between runaway demand and finite accelerator supply. Nvidia’s GPUs became the default currency of the AI boom because they were flexible, powerful, and supported by an unmatched software ecosystem.OpenAI’s problem is not that Nvidia makes bad chips. It is that Nvidia makes chips everyone else wants too. When the same scarce hardware is being chased by hyperscalers, model labs, sovereign AI projects, startups, research groups, and enterprises, even the best supplier relationship becomes a constraint.
Broadcom’s role is important precisely because it specializes in the less glamorous but deeply consequential world of custom silicon and high-scale networking. The AI race is not only about matrix multiplication. It is about moving data, keeping racks fed, cooling dense systems, and making clusters behave as reliable products rather than heroic engineering projects.
OpenAI’s earlier 10-gigawatt accelerator collaboration with Broadcom, announced in October 2025, already pointed in this direction. Jalapeño gives that partnership a named chip and a clearer narrative. The company wants custom accelerators, custom systems, and enough power-backed infrastructure to serve intelligence at industrial scale.
Inference Is Where AI Companies Either Print Money or Bleed It
The first wave of generative AI hype was obsessed with training frontier models. That made sense: bigger models, bigger clusters, and bigger benchmark jumps were the visible signs of progress. But the business model lives or dies on inference.Every ChatGPT conversation costs something. Every Codex request consumes compute. Every API customer that moves from demo to production converts AI from a capital-intensive bet into an operating-cost challenge. The more successful OpenAI becomes, the more punishing inference economics become unless the company can lower the cost per useful answer.
That is why performance per watt is such a loaded claim. OpenAI says early tests show Jalapeño delivering substantially better performance per watt than current state-of-the-art AI processors. The company has not provided enough public detail for outsiders to verify that claim independently, so it should be treated as vendor positioning until production deployments prove it.
Still, the direction is obvious. In a world where power availability is becoming as important as chip availability, a more efficient inference processor is not merely a nice engineering win. It is a way to stretch data center capacity, reduce serving costs, and make heavier AI products commercially viable.
The Apple Analogy Is Tempting Because It Is Partly Right
TechRadar and others have compared OpenAI’s move to Apple’s vertical integration playbook, and the analogy is useful if handled carefully. Apple’s advantage was not simply that it designed chips. It designed chips, operating systems, devices, APIs, and services as one coordinated product machine.OpenAI appears to be reaching for a version of that logic. The model architecture, inference runtime, product surface, API behavior, and chip design can all inform one another. If that loop works, OpenAI can optimize for the workloads it actually runs rather than for the average needs of the entire accelerator market.
But OpenAI is not Apple. It does not control the end-user device, the operating system, the browser, the enterprise desktop, or the smartphone distribution layer. Its “device” is the data center, and its customer experience is mediated through apps, APIs, cloud partners, and increasingly through other companies’ software.
That makes the integration play both more abstract and more fragile. Apple could use custom silicon to make a MacBook feel faster and last longer on battery. OpenAI has to use custom silicon to make remote intelligence feel faster, cheaper, and more reliable across millions of unpredictable workloads.
Broadcom Gives OpenAI a Shortcut, but Not a Free Pass
The most striking technical claim around Jalapeño is that the chip reportedly moved from initial design to manufacturing tape-out in nine months. If that timetable holds up under scrutiny, it is exceptionally fast for an advanced high-performance ASIC program. OpenAI has said its own AI models helped with aspects of the design process, which gives the story an appealing recursive twist: AI helped design the chip that will run AI.But chip development is full of cliffs hidden behind triumphant tape-out announcements. A working sample is not the same as reliable high-volume deployment. A benchmark does not equal fleet economics. A rack that runs well in a lab can behave differently when thousands are deployed under real customer load.
Broadcom reduces the risk because it knows the ASIC business, the networking layer, and the hyperscale supply chain. But even Broadcom cannot repeal physics, packaging constraints, thermal limits, yield realities, or the brutal calendar of semiconductor manufacturing.
That is why OpenAI’s language about a “multi-generation compute platform” matters. The first chip does not have to be perfect if it establishes the design loop. The real prize is learning how to build Jalapeño’s successors faster, more efficiently, and more tightly aligned with OpenAI’s model roadmap.
Microsoft Is in the Room Even When It Is Not Center Stage
For WindowsForum readers, the Microsoft angle is unavoidable. OpenAI’s infrastructure ambitions do not exist in a vacuum; Microsoft has been OpenAI’s defining cloud and commercial partner, even as OpenAI has diversified its infrastructure relationships. Azure remains a critical distribution and deployment channel for enterprise AI, and Microsoft’s own Copilot strategy depends on making AI cheap and reliable enough to embed everywhere.That creates an interesting tension. Microsoft has its own AI silicon work, including Azure Maia, and it has every reason to reduce dependency on any single accelerator vendor. OpenAI building custom inference chips could complement Microsoft’s ambitions if it increases available capacity and lowers costs for shared services. It could also complicate the stack if OpenAI’s preferred hardware path diverges from Azure’s broader infrastructure strategy.
The likely near-term outcome is coexistence rather than conflict. Hyperscale AI is becoming too large for a single chip family, a single cloud architecture, or a single procurement channel. Nvidia GPUs, AMD accelerators, cloud-native silicon, and custom ASICs can all coexist because demand is expanding faster than supply can rationalize itself.
Still, Jalapeño signals that OpenAI wants more leverage in the relationship. A model company that depends entirely on someone else’s cloud and someone else’s chips has limited bargaining power. A model company with its own accelerator roadmap becomes a more complicated partner.
The Enterprise Lesson Is Not to Buy the Hype, but to Watch the Unit Economics
Enterprise IT should resist the urge to treat Jalapeño as a procurement story today. No CIO is going to spec a Jalapeño server for the branch office. This is not a Windows workstation chip, a new x86 rival, or a general-purpose cloud instance family aimed at ordinary enterprise workloads.The practical relevance is upstream. If Jalapeño works, it could change the cost and latency profile of OpenAI-powered services. That eventually affects API pricing, availability, response times, throughput ceilings, and the kinds of AI features vendors can afford to bundle into productivity software.
For administrators and architects, the key question is not whether Jalapeño beats an Nvidia GPU in an isolated benchmark. The question is whether OpenAI can use custom inference hardware to make high-volume AI services more predictable. Enterprises do not merely need dazzling demos; they need service levels, compliance controls, capacity planning, and prices that survive budget season.
That is where custom silicon could matter most. If OpenAI can lower its marginal inference costs, it has more room to support longer context windows, heavier agent workflows, faster code generation, and richer multimodal experiences without turning every customer interaction into an expensive compute event.
The Name Is Cute; the Strategy Is Severe
“Jalapeño” is a playful name, but the strategy behind it is anything but whimsical. OpenAI is moving from the world of software iteration into an industry where mistakes cost enormous sums and timelines stretch across years. That shift changes the company’s risk profile.Software companies are used to shipping, measuring, and updating. Semiconductor programs demand earlier commitments and tolerate fewer late pivots. A chip designed around today’s inference assumptions may look brilliant if the model roadmap follows expectations, or awkward if a new architecture changes the bottlenecks.
That is the core strategic gamble. OpenAI’s confidence comes from its privileged view of its own workloads. It knows what users ask, how models respond, where latency accumulates, what kernels dominate, and what future products it wants to serve. Most chip vendors have to generalize across customers. OpenAI can specialize.
Specialization is powerful, but it narrows escape routes. If OpenAI’s future models require very different memory behavior, sparsity patterns, routing mechanisms, or runtime orchestration, the hardware roadmap has to keep up. The silicon game rewards conviction, but it punishes the wrong kind of certainty.
The Cloud Wars Are Becoming Power Wars
Cloud Wars framed Jalapeño as OpenAI entering the silicon game, and that is the right headline. But the deeper story is that the AI cloud is becoming a power and infrastructure war. Compute is no longer an abstract cloud resource that appears when a developer swipes a credit card.The October 2025 OpenAI-Broadcom announcement spoke in gigawatts, not merely chips. That language is revealing. AI infrastructure is now being planned at a scale where electricity, substations, cooling, land, supply chains, and network fabrics are strategic assets.
This is why the Stargate-style data center narrative matters. Whether every announced project arrives exactly as described is less important than the direction of travel. The leading AI companies are trying to secure multi-year access to the physical substrate of intelligence.
For users, that sounds remote until it isn’t. Capacity shortages become waitlists, throttling, higher prices, slower features, regional limits, or degraded service quality. Abundant inference capacity becomes better products. The chip is only one piece, but it is a piece OpenAI increasingly does not want to leave entirely to others.
Nvidia Is Still the Center of Gravity
None of this means Nvidia is suddenly in trouble. The company’s advantage is not just silicon performance; it is CUDA, developer familiarity, mature systems, networking, libraries, and a vast ecosystem of software and operational knowledge. Custom ASICs often win specific jobs while GPUs remain indispensable for experimentation, training, mixed workloads, and fast-changing research.OpenAI itself is unlikely to abandon GPUs. Frontier model development still benefits from flexible hardware and enormous existing tooling. Even if Jalapeño becomes a major inference engine for OpenAI’s own products, the company will likely continue using a blend of Nvidia, other accelerators, and partner infrastructure.
The more realistic threat to Nvidia is not a single chip. It is workload segmentation. As inference volumes grow, the most repetitive and economically sensitive workloads become attractive targets for custom silicon. GPUs keep the frontier moving; ASICs harvest the stable high-volume paths.
That is exactly how mature computing markets evolve. General-purpose hardware dominates early because flexibility matters most. As patterns stabilize and scale increases, specialized hardware starts carving out profitable lanes. Jalapeño is OpenAI’s claim that LLM inference has reached that stage.
Developers Will Feel This Indirectly Before They See It Directly
For developers building on OpenAI’s API, Jalapeño is unlikely to show up as a selectable backend with a spicy logo. Its impact will be mediated through product behavior. Faster responses, lower latency variance, more generous limits, lower prices, or new model capabilities would be the visible signs that the infrastructure bet is paying off.The same is true for Windows developers integrating AI into desktop apps, enterprise tools, and automation workflows. The chip does not change how a developer calls an API. It could change what those APIs can economically promise.
That distinction is important because AI infrastructure stories often get consumed as spectacle. The real developer question is mundane: can I build a feature on this service and trust that it will be fast, available, and affordable at scale? If custom inference hardware improves that answer, it matters even if developers never touch the silicon directly.
There is also a competitive angle. If OpenAI can lower the cost of serving its models, rivals may have to respond through their own silicon partnerships, cloud optimizations, or pricing changes. Developers benefit when infrastructure competition becomes product competition rather than merely benchmark theater.
The Corporate Story Is Getting Harder to Separate from the Technical One
Cloud Wars rightly notes that OpenAI’s hardware move lands amid ongoing speculation about the company’s long-term corporate structure and financial trajectory. Custom silicon makes OpenAI look less like an application company and more like an infrastructure company. That matters to investors, partners, regulators, and competitors.A company that controls more of its stack can argue that it has a durable moat. It can point to models, distribution, data center capacity, hardware design, and operational learning as mutually reinforcing assets. That is a more compelling story than “we rent GPUs and serve a chatbot,” especially if public-market ambitions ever become real.
But vertical integration also raises the stakes. OpenAI’s capital needs become larger. Its execution burden becomes heavier. Its exposure to supply chain risk, energy politics, and hardware depreciation becomes harder to ignore.
The company’s public narrative has long been about making advanced AI broadly available. Jalapeño gives that mission an industrial form. Broad access requires abundant compute; abundant compute requires hardware, power, and money at a scale few companies can assemble.
The First Jalapeño Harvest Will Test the Whole Theory
The crucial test is not whether OpenAI can unveil a chip. It is whether it can deploy Jalapeño into production workloads and make the user experience measurably better. Axios reported that OpenAI had begun testing the chip in lab settings for tasks similar to Codex queries, with plans to use it for customer queries later in 2026. That is the bridge from announcement to reality.Production inference will expose the hard parts. Can the chip handle real traffic patterns? Can OpenAI’s software stack route workloads intelligently? Can the systems run reliably at scale? Can the company measure enough savings to justify the engineering and capital commitment?
If the answer is yes, Jalapeño becomes more than a symbolic Nvidia hedge. It becomes the first step in a feedback loop where OpenAI’s models shape its chips and its chips shape its models. That would be a meaningful structural advantage.
If the answer is mixed, the company still learns. A first-generation custom accelerator can disappoint in some dimensions and still teach the organization how to build the next one. In silicon, the first product is often less important than the roadmap it enables.
The Spicy Chip Leaves a Practical Aftertaste
Jalapeño is not a consumer product, but it is one of the clearest signs yet that AI’s next phase will be decided below the application layer. The winners will not simply have the cleverest chatbot interface. They will have the most efficient path from power plant to prompt response.- OpenAI’s Jalapeño is designed for inference, the high-volume serving workload that determines whether AI products can scale economically.
- Broadcom’s role gives OpenAI access to custom ASIC, networking, and systems expertise that a software-first company cannot improvise overnight.
- The chip should be read as part of a broader 10-gigawatt infrastructure strategy rather than as a standalone component announcement.
- Nvidia remains central to AI infrastructure, but stable inference workloads are exactly where custom silicon can begin to peel away volume.
- Enterprise and developer impact will appear indirectly through pricing, latency, reliability, limits, and the ambition of AI features built on OpenAI services.
- The biggest risk is that hardware specialization locks in assumptions at a time when model architectures and agent workloads are still evolving quickly.
References
- Primary source: Cloud Wars
Published: Mon, 06 Jul 2026 15:00:00 GMT
Loading…
cloudwars.com - Related coverage: techradar.com
Loading…
www.techradar.com - Official source: openai.com
Loading…
openai.com - Related coverage: techcrunch.com
OpenAI unveils its first custom chip, built by Broadcom | TechCrunch
Named Jalapeño, the new processor was designed specifically for the unique needs of OpenAI's inference systems.techcrunch.com - Related coverage: tomshardware.com
Broadcom and OpenAI unveil custom-built Jalapeño inference processor — OpenAI's first chip is a massive reticle-sized ASIC built in an ultra-fast nine-month development cycle | Tom's Hardware
Firms fail to disclose (m)any details.www.tomshardware.com - Related coverage: techspot.com
Loading…
www.techspot.com
- Related coverage: investing.com
Loading…
www.investing.com - Related coverage: gadgets360.com
Loading…
www.gadgets360.com - Related coverage: axios.com
- Related coverage: pcgamer.com
When will the AI bubble burst? As OpenAI signs yet another megabucks deal with Broadcom, can anyone make sense of the trillions of dollars involved? | PC Gamer
A hundred billion here, a hundred billion there, pretty soon you're talking about real money.www.pcgamer.com - Related coverage: omni.se
Loading…
omni.se - Related coverage: techxplore.com
Loading…
techxplore.com - Related coverage: zeronoise.ai
Loading…
zeronoise.ai