Anthropic (ANT) Investment Research: How In-House Computing and Workflow Automation Affect Claude Revenue
As enterprises bring models in-house and automate routine workflows, this research examines Claude revenue, compute commitments and shareholder returns.
Enterprise Observations and Investment Research · November 22, 2026
Recent firsthand observations at the enterprise level show that a growing number of companies are building their own compute capacity, and some are also using AI to redesign fixed processes, coding well-defined steps into programs. Once a process goes live, routine work is executed directly by code, with no need to call a model for each item. This observation reflects how some enterprises are adopting the technology; its prevalence across the market remains to be quantified.
The two approaches affect Anthropic's revenue in different ways. When an enterprise moves inference onto its internal network, existing cloud API spending may convert into equipment and operating costs. When an enterprise goes further and codifies its rules into programs, demand for model services concentrates in development, revisions, and exception handling. Engineering teams may continue to renew Claude seats, but a rising number of program executions does not necessarily translate into a corresponding increase in usage-based revenue. The core of the investment assessment is whether the model revenue from new work can continue to exceed the reduction in call spending from mature processes.
Fixed-Process Automation and the Revenue Cycle of Model Services
Consider daily order reconciliation as an example. An engineer can ask Claude to identify field differences between two systems, sort out the order of processing for refunds and allowances, and generate matching code along with tests. After acceptance, a scheduler reads the database each day and verifies records one by one by order number, amount, and date; items that satisfy the rules are posted, and only exceptions are routed to manual handling. When input formats and rules are stable, normal items can be completed from start to finish with ordinary computation. What is actually saved is the labor time of having a person review every item, while model fees concentrate in building, revising, or investigating new exceptions.
This also defines the boundary of substitution. If the same order carries a dispute note written freely by the customer, understanding the text and establishing the context may still be worth paying inference fees for. The enterprise can leave this step to Claude, hand it to a local model, or assign it to staff. As a result, the same company may add development seats and build its own inference hosts while reducing external API usage for mature processes. Converting all "work completed through AI" uniformly into continuously growing token demand would overstate the market for model services.
Anthropic's revenue structure makes it especially sensitive to changes in usage volume. Of revenue for the first nine months, US$26.306 billion came from usage-based billing, accounting for 79.1%, and another US$6.936 billion came from subscriptions. Usage-based revenue also contributed approximately 79% of year-over-year incremental revenue. Over the same period, roughly 44% of revenue was routed through cloud marketplaces, yet marketplace customers accounted for approximately 52% of net receivables at period end. This gap may stem from payment terms. To determine whether growth is absorbing working capital, channel mix, receivables aging, and actual collections must be examined together. Public data remain insufficient to break out realized pricing, user conversion paths, or product-level gross margin for Claude Code, Cowork, and pure API. SEC Prospectus, Revenue Breakdown and Customer Concentration
Nine-month data show that revenue growth has already lowered compute costs as a share of revenue. Revenue for the first nine months reached 13.1 times the prior-year period, compute and infrastructure expenses fell from approximately 172% of revenue to 57%, and operating cash flow reached US$6.569 billion. Third-quarter cost trends, however, reveal profit pressure. Compared with the second quarter, third-quarter revenue increased by US$5.401 billion, while compute and infrastructure expenses rose by US$4.305 billion over the same period, an incremental ratio of 79.7%. The company used approximately US$5 billion of short-term compute during the quarter. How much of this represents a premium relative to longer-term contracts has not been disclosed. The company's adjusted operating income rose from US$2.235 billion to US$2.463 billion, yet the margin fell from 19% to 15%. These expenses include both training and inference, so the 79.7% figure cannot be applied directly to the gross margin on the next customer. SEC Prospectus, Quarterly Income Statement and Adjusted Operating Income
Hardware Prices and the Cost of On-Premises Enterprise Deployment
Recent firsthand observations from Asian distribution channels show that spot quotes for some SSDs and RTX 5090 cards have fallen to 70–80% of earlier levels. Quotes for the same models in Akihabara, Japan, show similar declines. On November 7, ITmedia reported that a restocked Palit RTX 5090 GameRock at Dospara's Akihabara flagship store was priced at ¥1,288,000. On November 13, the in-store price was ¥999,800, a one-week decline of 22.4%, or about 77.6% of the prior restock price (a 7.76 ratio, or "7.76 zhe"), and store staff confirmed that units had sold after the price cut. For SSDs, AKIBA PC Hotline!'s survey for the first half of October recorded the WD Black SN8100 2TB falling from ¥91,440 in the prior survey to ¥70,980, a decline of 22.4%, likewise about 77.6% of the prior price (7.76 zhe). These figures reflect prices for specific models and channels; average prices across Asia and enterprise-grade procurement contracts still need to be observed separately. RTX 5090 restock price · In-store check after the price cut · SSD model and price survey
For an enterprise that already has a host machine and needs only one graphics card to keep sensitive data on its internal network, a single-card price cut can directly change the purchasing decision. For building a machine from scratch, only the cost of the discounted components declines. For example, if the GPU and SSD together account for 60% of the total system and both fall 22.4% while other prices are unchanged, total system cost declines by roughly 13.4%. This is a different scenario from the valuation of existing equipment at 30% of original price discussed below, and the two discounts are not applied cumulatively. The impact on ANT revenue ultimately depends on how much of the work that would otherwise have been paid to Claude is absorbed by the new equipment. Processes that have been codified and no longer use a model at execution do not need to wait for hardware prices to fall.
Breakeven Analysis Assuming Existing Equipment Valued at 30% of Original Price
The economic cost of occupying existing equipment is set at 30% of its original price, amortized evenly over the remaining three years with zero residual value at the end. For equipment that is fully paid for and idle, this amortization represents the forgone value of resale or alternative use, and there is no equivalent new cash outlay. Two configurations with verifiable pricing are used: four existing M4 Pro Mac minis, based on Apple Taiwan's launch starting price of NT$46,900 each, and one NVIDIA DGX Spark, based on the official announced price of US$4,699 converted at a scenario exchange rate of 32. Macs with upgraded memory should be recalculated using the original invoice. The two are compared only on an economic threshold; performance must be validated task by task. Apple historical pricing · NVIDIA price announcement · DGX Spark specifications
Scenario; all amounts in New Taiwan dollars | Four Mac mini M4 Pro | One NVIDIA DGX Spark |
|---|---|---|
Value of full set of equipment at 30% of original price | NT$56,280 | NT$45,110 |
Annual equipment amortization | NT$18,760 | NT$15,037 |
Assumed average power of the full set, year-round | 200 W | 150 W |
Annual electricity cost, assuming 8,760 hours at NT$4 per kWh | NT$7,008 | NT$5,256 |
Annual maintenance labor, 2 hours per month at NT$1,000 per hour | NT$24,000 | NT$24,000 |
Total annual fixed cost | NT$49,768 | NT$44,293 |
Qualified items required to break even, rounded up | 10,369 items/year | 9,228 items/year |
Year-round operation, minimum average per day | 29 items | 26 items |
Both cases use one piece of work that passes the same quality and latency acceptance as the unit. The scenario API fee is US$0.20 per item (NT$6.40), and the local variable processing cost is NT$1.60 per item, excluding the data center electricity and maintenance labor already listed, for a net saving of NT$4.80 per item. The breakeven item count equals annual fixed cost divided by net saving per item. If traffic exists on only 250 working days, the two cases must reach 42 and 37 items per working day, respectively. Power, electricity price, labor hours, and unit prices are all scenario values, and existing network and space are assumed to be reused. Additional connectivity, redundancy, build-out, or manual rework must be counted separately. If the API price falls a further 25% at the same time, net saving per item shrinks to NT$3.20, and the daily year-round threshold rises to 43 and 38 items. If the actual API price is already below the local per-item variable cost, no amount of additional usage can break even by amortizing the equipment.
Multiple Macs can already jointly run inference on a single large model. Apple's MLX can shard a model across multiple Macs, using each machine's memory and compute so that open-weight models too large for a single machine can run on an enterprise's internal network. Beginning with macOS 26.2, Apple silicon Macs with Thunderbolt 5 support remote direct memory access (RDMA) over Thunderbolt, reducing inter-node communication latency. Apple's technical documentation demonstrates a cluster connection using four M4 Pro Mac minis. WWDC26 separately demonstrated distributed inference of the one-trillion-parameter Kimi K2.6 model on four Mac Studios. The two examples differ in machine model, memory, and validation items. Apple Technical Note TN3205 · WWDC26 large-model demonstration, at the 82-minute mark · Apple MLX distributed inference research
For Anthropic, this capability widens the range of tasks that could be moved back onto an internal network. An enterprise that already has four Mac minis can run independent small tasks across separate machines, or have all four jointly serve a larger model. The former mainly increases the number of items that can be processed concurrently, while the latter gives the enterprise an opportunity to complete locally work that was previously sent externally because a single machine lacked capacity. The DGX Spark offers a single-machine option with 128 GB of unified memory. The 29 and 26 items per day in the preceding table are breakeven thresholds under the specified cost assumptions. Which configuration can reach that threshold with less rework depends on the actual model, quantization precision, context, concurrency, and acceptance criteria. This research has not obtained same-task benchmarks for the two sets of equipment. WWDC26 local agents and MLX distributed inference · NVIDIA DGX Spark specifications
The daily breakeven item counts in the preceding table apply only to work that still requires a model at execution. If a rule can already pass acceptance reliably, the enterprise should compare the costs of development, testing, maintenance, and ordinary computation. Purchasing an inference host for it would instead overstate the capital required for automation. For lean teams, first coding stable rules into programs and then letting a small number of complex cases be handled by local or cloud models allows the three approaches to coexist within the same process.
Global AI Spending and the Revenue Available to Anthropic
Working backward from annual revenue of US$197.5 billion in the middle scenario of the valuation below, if Anthropic could capture 30% of the serviceable market, the market recognizable by the same tier of suppliers would need to reach US$658.3 billion. If it could capture only 20%, the threshold rises to US$987.5 billion. Gartner forecasts global AI spending of US$4.05 trillion in 2027, of which infrastructure accounts for US$2.397 trillion, generative models approximately US$51 billion, agents and assistants approximately US$67 billion, and AI software approximately US$649 billion. These categories do not align with the company's three-year scenario year, and software subscriptions and the API fees within them cannot be double-counted. What bulls must demonstrate is that Claude can continuously convert budgets previously paid to enterprise software, professional services, and process vendors into revenue it can recognize itself. Growth in global AI spending alone cannot confirm that it will capture the required share. Gartner forecast of global AI spending by category
Open-weight models give enterprises another option: assign routine work to cheaper models and call Claude only for difficult steps. OpenRouter's research shows that Chinese open models approached 30% of total token volume on its platform in some weeks. This platform aggregates buyers who switch between models, and usage is also affected by unit price, so the figure cannot be extrapolated to global revenue share. A counterexample favorable to Anthropic is that inference prices are higher but save enough manual review. Assuming each item costs an additional US$0.50 and review labor costs US$60 per hour, a more expensive model could save the enterprise money if it saves more than 30 seconds per item on average. Demonstrating that such a premium is durable requires acceptance results, actual fees paid, and rework time for the same batch of tasks. OpenRouter usage study · DeepSeek official model card
The relative strength of US and Chinese AI affects where Anthropic can sell and how many alternatives customers have. Stanford reports that US private AI investment in 2025 was US$285.9 billion, versus US$12.4 billion for China. Government-guided funds mean this comparison understates China's total investment. US capital and cloud channels help in obtaining compute, but weights released by Chinese teams can be deployed in third countries, depressing the price local customers are willing to pay Claude. Anthropic's sales policy also restricts certain overseas organizations controlled from unsupported regions. Even if no API fee is received from a given transaction, Chinese models may still affect ANT's pricing. Even if the US industry maintains its lead, the gains must still be distributed among chip, cloud, model, and application suppliers. Stanford AI Index · Anthropic sales region policy
Compute Payment Obligations When Demand Declines
Observation period | Company disclosure | Conclusion that can be drawn |
|---|---|---|
Twelve months from end of September 2026 | Expected payments of US$70.7 billion, of which US$4.2 billion is subject to specified conditions or termination rights before performance begins | The remaining US$66.5 billion accounts for 94.1%. A customer's own reduction in usage is not a termination condition of the type listed by the company |
Full-term contractual obligations | Total of US$900.9 billion, of which US$186.2 billion is subject to the specified categories above | The remaining US$714.7 billion accounts for 79.3%, spread across multiple years. Delivery, performance, and other remedy provisions may still affect the timing or amount actually paid |
Three cloud providers with explicit minimum purchase commitments | Microsoft US$47.2 billion, Google US$121.3 billion, Amazon US$110.0 billion | Full-term minimum purchases across the three contracts total US$278.5 billion. Google and Amazon explicitly list shortfall payments and exceptions. Some contract periods have already passed |
The above ranges overlap. The prospectus does not disclose the precise proportion that all contracts combined would still require paying at zero usage. The 94.1% figure is the complement of a specific contract category. The full-term minimums of the three cloud providers also lack unpaid balances as of the same date, so dividing US$278.5 billion directly by US$900.9 billion cannot estimate a "non-cancelable ratio." Contract flexibility does differ; for example, after SpaceXAI's initial non-cancelable period ends, the parties may terminate early on 90 days' notice. Another arrangement with no specified fixed amount carries company-estimated spending of US$150 billion to US$175 billion, with the actual amount varying with capacity, delivery, and price. SEC Prospectus, pages 178–180, Contractual Obligations
Short-term liquidity depends on payment timing and available funds. The US$70.7 billion expected to be paid within one year equals 28.9% of pro forma cash and short-term investments of US$244.6 billion after full exercise of the greenshoe. The company has time to adjust. Actual cash after closing, future collections, other expenditures, and financing conditions determine how long that period lasts. Operating cash flow for the first nine months less equipment purchases and related prepayments gives a simplified balance of US$3.689 billion, which is also affected by working capital timing. If demand slows, shareholders may first bear margin compression and dilution from equity financing. The problem would escalate into an existential crisis only if low utilization persisted while cash burn and refinancing pressure rose together. Using fixed-price capacity as an illustration, if utilization falls from 80% to 60%, the fixed cost allocated to each unit of actual usage rises by one-third. The company's actual capacity utilization has not been disclosed. SEC capitalization table, cash flow statement, and contract notes
Revenue and Cash Flow Conditions Implied by the US$88.91 Share Price
The November 20 closing price of US$88.91 is 7.1% above the US$83 offering price and 8.5% below the first-day closing price. Based on approximately 12.5198 billion pro forma shares outstanding after full exercise of the greenshoe, basic equity value is approximately US$1.113 trillion, or 31.5 times trailing twelve-month revenue, or 16.4 times annualized third-quarter revenue. The latter multiplies a single quarter by four and is used only to show the effect of the denominator on the multiple. Approximately 2.848 billion additional options and RSUs are not included in this basic share count, and actual dilution still depends on exercise, vesting, settlement, and other conversions. Company IPO closing announcement · SEC capitalization table · Historical closing prices
Assume investors require a 10% annualized return, no cash is distributed over three years, and the share count is unchanged. Equity value at the end of the period would then need to reach approximately US$1.482 trillion. If the market then assigns 30 times equity free cash flow, annual equity free cash flow would need to approach US$49.4 billion. If 25% of revenue can be converted into this cash flow, annual revenue would need to reach approximately US$197.5 billion. Starting from annualized third-quarter revenue of US$67.824 billion, this would require compound growth of approximately 43% for three consecutive years. If it is not reached until the fifth year, then under the same return requirement, multiple, and cash flow rate, the required annual revenue rises to approximately US$239.0 billion. The above is a valuation back-solve and does not separately account for dilution during the period or excess cash distributions.
Same year-three revenue of US$197.5 billion achieved | Equity free cash flow rate | Ending multiple | Estimated annualized return from buying at US$88.91 |
|---|---|---|---|
Revenue grows and cash conversion also meets target | 25% | 30x | About 10.0% |
Legacy compute, equipment, and other expenses erode cash | 20% | 30x | About 2.1% |
Weaker cash conversion, and the market also lowers the multiple | 20% | 25x | About −3.9% |
Even if revenue three years from now reaches 2.9 times the annualized third-quarter level, shareholder returns could still be negative. The current price places demands on revenue growth, equity free cash flow rate, and ending valuation multiple alike. Of the three, the cash flow rate depends in particular on the utilization efficiency of committed compute.
Dot-Com Bubble Lessons and Shareholder Returns
Global Crossing went through bankruptcy reorganization, yet the fiber it laid remained available to later enterprises and customers. AI may likewise spread across more industries, and new application-layer companies may also earn profits on cheaper hardware, while first movers that locked in high-cost capacity early bear the losses. Anthropic already has paying demand and a pro forma cash buffer, so there is no evidence that it is "about to collapse." A nearer-term risk, and one more readily verifiable through financial reports, is that it must sustain R&D and honor compute purchases yet may be unable to sell all of its capacity at adequate prices. FCC document on Global Crossing's reorganization and asset transfer
What to track next is the same batch of processes that have gone live. Each quarter, record separately how many items they complete, how many run on code alone, how many use local models, and how many still pay to call Claude, and then separately track the seat and API revenue generated by new process development, revisions, and maintenance. Processes that previously relied on manual work and were coded with AI for the first time create new value. Processes that shift from calling Claude item by item to executing code directly are the ones that cut existing usage-based bills directly. If the two are mixed together, revenue not yet captured is easily misrecorded as churn.
Evidence supporting an upward revision of valuation includes combined revenue from new process development, complex tasks, and maintenance exceeding the reduction in call spending from mature processes, with customer collections and per-share cash generation improving in tandem. Conversely, if the volume of work completed by the same customers increases while the total paid to Claude continues to decline, and Anthropic's minimum purchase payments keep rising, the idle-capacity risk that shareholders must bear will increase. The company has disclosed revenue and supply commitments but has not provided the customer process-level data described above. Measured against the US$88.91 closing price, the available evidence remains insufficient to support the view that the valuation carries an adequate margin of safety.
