These figures, reported by The Information from documents shared with shareholders, underscore the central challenge facing the AI industry: explosive demand is not yet translating into sustainable profits . The stakes are high, as both OpenAI and Anthropic have reportedly filed secret IPO applications in mid-2026, forcing them to show a credible path to profitability .

The Software Fix: Multi-Model Routing

The specific cost-cutting software optimization OpenAI has developed to reduce inference costs by more than half is multi-model routing. This technique automatically routes simpler queries to smaller, cheaper models while reserving the most powerful (and expensive) models only for complex requests .

Industry guides describe the hybrid approach as delivering 30-75% cost reduction at the model layer by using quantization, distillation, and right-sizing, combined with runtime optimizations like batching, speculative decoding, and KV cache management that can yield 40-80% throughput gains . OpenAI has also operationalized this by offering 50% price cuts for deferred inference via its Batch API—if customers agree to wait up to 24 hours for results .

Beyond routing, engineers at OpenAI have told The Information that the techniques are powerful enough that at one point they served ChatGPT to logged-out visitors on just a couple hundred Nvidia GPUs .

The Hardware Initiative: Jalapeño Custom Inference Chip

On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-built inference chip . Key details:

• Type: Application-specific integrated circuit (ASIC)—not a GPU, purpose-built solely for LLM inference
• Cost reduction: Early testing shows ~50% lower cost per inference compared to current NVIDIA GPU-based inference hardware
• Performance: Delivers "substantially better" performance per watt than current state-of-the-art
• Development: Designed from scratch to production in just nine months, accelerated by OpenAI's own AI models assisting in chip design
• Deployment: Gigawatt-scale deployment targeted for end of 2026
• Long-term commitment: Part of a 10 GW infrastructure commitment through 2029
• Annual savings: The 50% cost reduction could translate to $500M–$1B in annual savings for OpenAI by 2028

OpenAI describes the chip as "the first building block in a multi-year custom silicon roadmap" to build its own computational stack and address NVIDIA's pricing power . The chip is vertically integrated—OpenAI will use it exclusively for its own inference workloads, not sold to external customers .

Financial Pressures in Detail

Key figures from the Q1 2026 shareholder documents:

• Revenue: $5.7 billion (roughly tripled year-over-year)
• Cash burn: $3.7 billion (also tripled year-over-year)
• Operating loss: $9.3 billion in Q1 alone
• Net loss: $21.3 billion (including stock-based compensation of $2.3 billion and a $12.4 billion non-cash accounting charge)
• R&D expenses: $8.6 billion
• Gross margin: Improved slightly to 39%
• Cash reserves: $73 billion in cash and securities, providing runway despite massive losses

The company's operating margin was roughly -122%, meaning it lost about $1.22 for every dollar earned . In 2025 overall, OpenAI posted approximately $38.5 billion in losses on $34 billion in spending .

The Broader Industry Race: Inference Costs Collapsing

OpenAI's cost-cutting moves fit into a broader industry-wide price war that is reshaping the economics of AI:

• Anthropic has accelerated customer growth and surpassed OpenAI in U.S. business adoption (34.4% vs 32.3%) as of May 2026 .
• Open-source models are compressing pricing: four Chinese open-weights models now match frontier coding benchmarks at a third of the cost of proprietary alternatives . DeepSeek V4 was reported to be fifty times cheaper to run than Claude Opus 4.6 .
• Google's Gemini 3.1 Flash costs $0.10 per million input tokens—a 99.7% price reduction in three years .
• Anthropic's Claude Sonnet 4.6 outperforms GPT-4 on every legacy benchmark at $3 per million input tokens .
• Inference costs have fallen roughly 50x in three years, with the market projected to grow from $103 billion (2025) to $255 billion (2030) .
• OpenAI is considering significant per-token price cuts for developers and enterprise customers to preempt similar moves by Anthropic .
• Customers are increasingly using multi-model routing tools to mix and match cheaper models and avoid premium pricing from any single vendor .

What This Means for the AI Industry

The combination of software and hardware optimizations positions OpenAI to potentially halve its inference costs even as it scales. But the company is not acting in a vacuum. The broader collapse in inference pricing—driven by competitors, open-source alternatives, and specialized hardware—means that pricing power is eroding across the industry .

For enterprises, this is good news: the effective cost of running AI workloads is falling faster than capabilities are growing. OpenAI's Batch API already offers 50% discounts for deferred workloads, while multi-model routing can cut costs by 40-60% without sacrificing quality . The question is whether any single provider can maintain premium pricing in a market where Chinese open-source models offer frontier-level capability at a fraction of the cost .

These figures, reported by The Information from documents shared with shareholders, underscore the central challenge facing the AI industry: explosive demand is not yet translating into sustainable profits . The stakes are high, as both OpenAI and Anthropic have reportedly filed secret IPO applications in mid-2026, forcing them to show a credible path to profitability .

The Software Fix: Multi-Model Routing

The specific cost-cutting software optimization OpenAI has developed to reduce inference costs by more than half is multi-model routing. This technique automatically routes simpler queries to smaller, cheaper models while reserving the most powerful (and expensive) models only for complex requests .

Industry guides describe the hybrid approach as delivering 30-75% cost reduction at the model layer by using quantization, distillation, and right-sizing, combined with runtime optimizations like batching, speculative decoding, and KV cache management that can yield 40-80% throughput gains . OpenAI has also operationalized this by offering 50% price cuts for deferred inference via its Batch API—if customers agree to wait up to 24 hours for results .

Beyond routing, engineers at OpenAI have told The Information that the techniques are powerful enough that at one point they served ChatGPT to logged-out visitors on just a couple hundred Nvidia GPUs .

The Hardware Initiative: Jalapeño Custom Inference Chip

On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-built inference chip . Key details:

• Type: Application-specific integrated circuit (ASIC)—not a GPU, purpose-built solely for LLM inference
• Cost reduction: Early testing shows ~50% lower cost per inference compared to current NVIDIA GPU-based inference hardware
• Performance: Delivers "substantially better" performance per watt than current state-of-the-art
• Development: Designed from scratch to production in just nine months, accelerated by OpenAI's own AI models assisting in chip design
• Deployment: Gigawatt-scale deployment targeted for end of 2026
• Long-term commitment: Part of a 10 GW infrastructure commitment through 2029
• Annual savings: The 50% cost reduction could translate to $500M–$1B in annual savings for OpenAI by 2028

OpenAI describes the chip as "the first building block in a multi-year custom silicon roadmap" to build its own computational stack and address NVIDIA's pricing power . The chip is vertically integrated—OpenAI will use it exclusively for its own inference workloads, not sold to external customers .

Financial Pressures in Detail

Key figures from the Q1 2026 shareholder documents:

• Revenue: $5.7 billion (roughly tripled year-over-year)
• Cash burn: $3.7 billion (also tripled year-over-year)
• Operating loss: $9.3 billion in Q1 alone
• Net loss: $21.3 billion (including stock-based compensation of $2.3 billion and a $12.4 billion non-cash accounting charge)
• R&D expenses: $8.6 billion
• Gross margin: Improved slightly to 39%
• Cash reserves: $73 billion in cash and securities, providing runway despite massive losses

The company's operating margin was roughly -122%, meaning it lost about $1.22 for every dollar earned . In 2025 overall, OpenAI posted approximately $38.5 billion in losses on $34 billion in spending .

The Broader Industry Race: Inference Costs Collapsing

OpenAI's cost-cutting moves fit into a broader industry-wide price war that is reshaping the economics of AI:

• Anthropic has accelerated customer growth and surpassed OpenAI in U.S. business adoption (34.4% vs 32.3%) as of May 2026 .
• Open-source models are compressing pricing: four Chinese open-weights models now match frontier coding benchmarks at a third of the cost of proprietary alternatives . DeepSeek V4 was reported to be fifty times cheaper to run than Claude Opus 4.6 .
• Google's Gemini 3.1 Flash costs $0.10 per million input tokens—a 99.7% price reduction in three years .
• Anthropic's Claude Sonnet 4.6 outperforms GPT-4 on every legacy benchmark at $3 per million input tokens .
• Inference costs have fallen roughly 50x in three years, with the market projected to grow from $103 billion (2025) to $255 billion (2030) .
• OpenAI is considering significant per-token price cuts for developers and enterprise customers to preempt similar moves by Anthropic .
• Customers are increasingly using multi-model routing tools to mix and match cheaper models and avoid premium pricing from any single vendor .

What This Means for the AI Industry

The combination of software and hardware optimizations positions OpenAI to potentially halve its inference costs even as it scales. But the company is not acting in a vacuum. The broader collapse in inference pricing—driven by competitors, open-source alternatives, and specialized hardware—means that pricing power is eroding across the industry .

For enterprises, this is good news: the effective cost of running AI workloads is falling faster than capabilities are growing. OpenAI's Batch API already offers 50% discounts for deferred workloads, while multi-model routing can cut costs by 40-60% without sacrificing quality . The question is whether any single provider can maintain premium pricing in a market where Chinese open-source models offer frontier-level capability at a fraction of the cost .