[TL;DR Daily Highlights] What I’m watching isn’t how expensive GPUs are—but whether they can consistently produce billable “intelligence” like a proper intelligent agent factory. DeepSeek’s disclosure of $38 per GPU hour is like lifting a corner of the AI industry’s most sensitive ledger: the question is no longer just how much money is being burned, but whether each GPU can pay for itself. 📌 First, the most eye-catching figures: • DeepSeek disclosed: a single cluster can process 15.5 trillion tokens per day. • At publicly listed off-peak API pricing, this translates to roughly $38 in theoretical revenue per GPU hour. • Extrapolated annually, a single GPU could generate over $330,000 in billable output. This isn’t profit—or even easy money—it’s merely theoretical billing capacity. But it’s critical: it shows that if inference optimization is pushed hard enough, GPUs need not remain cost black holes; they can become machines that generate cash flow. 📌 How was $38 squeezed out? • DeepSeek V4.1 Flash has 552B total parameters, using a MoE architecture. • During input processing, ~8B parameters are activated; during output generation, ~16B. • Global KV Cache usage reduced to ~1/4 of the previous generation; long-term cache reduced to ~1/8. • Each token’s global KV Cache is ~890 bytes, supporting up to 1M context tokens. In other words, it’s not about using more expensive hardware—it’s about ruthless systems engineering: expert parallelism, speculative decoding, asynchronous scheduling, separating prefill and decode phases, and extending KV Cache retention. Simply put: among the same pool of GPUs, whoever minimizes redundant computation comes closest to printing money. 📌 But don’t treat $38 as an industry average yet. • DeepInfer hasn’t fully open-sourced all production components. • Critical elements like KV Cache systems, service orchestration, and cross-replica load balancing still require external integration. • It’s currently optimized primarily for DeepSeek-V4.1 Flash; other models can’t be directly replicated. Downloading the framework ≠ replicating revenue. $38 is more like a “showroom model under peak engineering conditions”—not a wealth formula anyone can copy with a few GPUs. 📌 Why are so many model companies still losing money despite stronger inference? • Stanford HAI / Epoch AI estimate: OpenAI’s 2025 compute spend will be ~$16.3B, with $8.3B for R&D and $8.0B for inference. • Compared to 2024, OpenAI’s inference spend rose from $1.8B to $8.0B—roughly a 3.4x increase. • Anthropic’s 2025 compute spend is estimated at ~$6.8B: $4.1B for R&D, $2.7B for inference. Here’s the truth: training and inference are two separate ledgers. You can sell tokens, but that doesn’t mean you stop burning cash to build the next model—and as users grow, so does your inference bill. Lower unit cost doesn’t mean lower total cost. The AI industry is like building highways: more cars mean more toll revenue—but also higher maintenance, expansion, and interest costs. 📌 Can you build a viable business selling just tokens? Answer: Possibly—but it won’t be easy. The article presents a straightforward sensitivity analysis: • If full cost burden is $10/GPU hour; • Theoretical peak billing capacity is $38/GPU hour; • Then achieving just 26.3% effective billing utilization would cover costs. • At 50% effective billing, hourly revenue would be ~$19, leaving $9 profit after $10 cost. This model is alluring—but reality isn’t romantic. Throughput, concurrency, latency, customer mix, peak traffic, and idle time all erode projected margins. And when price wars hit, much of the efficiency gain often gets passed on to customers. 📌 CoreWeave shows the other side very clearly: • Q2 2026 revenue: $2.575B (+112% YoY). • Adjusted EBITDA: $1.51B (59% margin). • But GAAP net loss: $626M. • Depreciation & amortization: $1.393B; net interest expense: $640M. See? The compute business isn’t “revenue = profit.” GPUs depreciate. Growth is funded by capital. Interest follows you like a shadow. Gross margin doesn’t guarantee net profitability. 📌 The application layer is even more awkward: demand rises, but profitability doesn’t always follow. • Cursor’s gross margin hit -23% in the January 2026 quarter; if free-user inference costs are fully counted, it drops to -31%. • https://t.co/RxWvQzBAAu filed for bankruptcy protection with ~12M monthly active users and $5.3M revenue in August 2026—but still lost $8–9M per month. • In just the first eight months of 2026, https://t.co/RxWvQzBAAu spent $56.1M on compute—45% of total operating expenses. • Meanwhile, Midjourney generated over $200M in revenue in 2023—and remained profitable. Same AI-native approach, wildly different outcomes. The reason isn’t mysterious: using AI doesn’t automatically create a viable business model—the key is whether the value delivered justifies the inference and operational costs. 📌 The bigger opportunity may lie in Agents: • Menlo Ventures estimates global consumer AI spending reached $40B in 2026—more than triple the $12B in 2025. • Yet U.S. adult AI adoption rose only from 61% to 64%. In other words: new users didn’t explode—existing users simply used AI more deeply and paid more. The market is now betting not just on chat volume, but on how many daily tasks AI can complete for you: emails, procurement, scheduling, software operations, cross-service coordination—these are where Agents could truly amplify token consumption. But I must pour cold water: more tokens don’t equal better business. If an Agent product earns $20/month but spends $30 on models and tools, it’s just subsidizing upstream GPU owners—not sustaining itself. My conclusion is simple: 1. $38/GPU hour shows AI inference *can* be viable—and the value of engineering optimization is severely underestimated. 2. But from “billable” to “profitable,” there are four walls to cross: utilization, pricing power, capital cost, and product value. 3. The companies that truly survive cycles won’t necessarily be the ones burning the most GPUs—but those who turn tokens into real-world task outcomes.If, in the future, every agent works, shops, negotiates, and executes processes on behalf of humans, the question will no longer be “How much does the token sell for?” but rather—how much value has the intelligence I created generated for you? This is the account I care about most.