Convergence: Part 3: AI
Part 3: AI
OpenAI and Anthropic
In 2025, OpenAI reached an annualized revenue run rate (ARR) >$20 billion
Anthropic reached an ARR of $19 bn by early 2026.
Yet, these figures mask severe cash burn rates, compressing gross margins, and immense infrastructural debts that challenge the sustainability of the current subscription models.
To answer the central question - how much OpenAI and Anthropic would need to charge in monthly subscription revenue per account to achieve strict, unsubsidized profitability - it is necessary to construct a holistic, zero-based cost model. This model must
eliminate external capital life support
apply market-rate cloud computing costs
allocate multi-billion-dollar training amortizations directly to the active paying user base
factor in the soaring prices of
specialized engineering talent
revenue-sharing agreements
intellectual property licensing.
Based on the synthesis of 2025 and 2026 financial leaks, compute cluster pricing, hardware efficiency scaling, and active user volume data, the analysis indicates that the era of the $20-per-month flat-rate AI subscription is mathematically unsustainable in a free-market vacuum.
To achieve break-even profitability without external venture subsidies or free-tier cross-subsidization:
OpenAI would require a blended monthly subscription rate of approximately $177 per active premium account.
Anthropic, possessing a developer-heavy enterprise user base engaging in token-intensive agentic coding workflows, would dictate a baseline subscription requirement closer to $250 per month
When accounting for debt-servicing costs and multi-year infrastructural lock-ins (OpenAI's $1.4 trillion data center and Anthropic's multi-bn dollar cloud revenue-sharing deals) - the unsubsidized break-even requirement spikes to approximately
$356.26USD per month for OpenAI and
$430.55USD per month for Anthropic
Assuming the companies’ own projections of user increases materialize, budgets do not over-run, energy crises do not eventuate, interest rate environments stay current, and the companies reach profitability as anticipated (2030 and 2029 respectively)
The Illusion of Software Margins
The traditional Software-as-a-Service (SaaS) financial model is characterized by
High upfront r&d costs
Near-zero marginal costs of distribution
Frontier AI companies have fundamentally decoupled from this economic reality.
GenAI models possess
Heavy unavoidable marginal cost for every interaction (inference cost)
Recurring CapEx required to train successive generations of models.
In 2024, OpenAI set an adj. gross margin of 46%. By 2025, this figure dropped to 33%.
Driven by necessity to purchase expensive, short-notice computing capacity to meet unexpected spikes in global demand, causing inference costs to quadruple.
The underlying driver of this margin collapse is the shifting nature of AI interactions. The evolution from single-turn conversational queries to complex, multi-step "agentic" reasoning requires exponentially more compute per user session.
For instance, OpenAI's GPT-4o had a 128k-token context window. GPT-5 supports up to a 400k-token window (See AI Tokens and Pricing)
The scale of capital required to sustain this architecture is historic and largely unprecedented in the private sector.
OpenAI has raised its spending projections to a cumulative cash burn of $665 bn by 2030, with the company not expected to become cash-flow positive until then.
Anthropic expected a cash burn of $5.2 bn against $9 bn in projected revenue for 2025 (Muppidi, 2026)
Anthropic acknowledged its inference costs (Google and Amazon servers) were 23% higher than the company expected
The $20 monthly subscription fee currently charged by the platforms is an artificial construct designed for market capture.
It bears little relation to the underlying cost of goods sold and acts as a barrier to understanding the true unit economics of AI.
Macro Expense Architecture
In an unsubsidized model, there is no VC to absorb losses, nor are there hypothetical revenues to float the consumer base. The costs must be borne entirely by the users generating the demand.
These expenses fall into four pillars:
Hardware, Datacenter, and Energy Economics;
Model Training and Capital Expenditure;
Talent and Human Capital Liabilities;
Selling, General, and Administrative (SG&A) Overhead, including intellectual property licensing and corporate revenue sharing.
Pillar 1: Hardware, Datacenter Buildouts, and Energy Consumption
The foundational layer of AI unit economics is silicon hardware and raw electricity.
GenAI operates on vast, globally distributed clusters of specialized accelerators, primarily Nvidia H100 GPUs and Google Tensor Processing Units (TPUs).
To calculate unsubsidized costs, one must look beyond the heavily discounted compute credits provided by Microsoft Azure or Google Cloud to their respective partners, and instead examine the open market rates for bare-metal compute access.
Throughout 2024 and 2025, the cloud compute market experienced intense volatility.
During peak scarcity in late 2024, H100 instances were priced between $8-$10/hour.
By late 2025 and early 2026, accelerated datacenter buildouts brought the baseline on-demand rate down significantly.
Major providers like AWS offered P5 instances at roughly $3.50-$3.90/hour
Long-term enterprise commitments could be as low as $1.90-$2.10/hour.
Even with these price reductions, which represent a 64% to 75% decline from historical peaks , the volume of active clusters required to serve global inference demands results in high daily operational expenses.
This hardware requires immense electrical power.
OpenAI’s compute capacity expanded nearly tenfold over a two-year period, reaching 1.9 gigawatts in 2025.
To contextualize this, 1.9 gigawatts is equivalent to the peak electricity consumption of roughly two million modern households.
Standard query on GPT-4o model consumes ~0.3 watt-hours of electricity
Equivalent query utilizing GPT-5 architecture consumes an ~18 watt-hours.
This represents a 60-fold increase in pure energy expenditure per query for advanced reasoning tasks.
In late 2025, Anthropic expanded its partnership with Google Cloud, committing to a multi-year deal worth tens of billions of dollars to access up to one million TPUs, bringing well over a gigawatt of capacity online by 2026.
When cloud infrastructure costs are modeled without equity-based subsidies, Anthropic projections expect to incur an estimated $80 bn in cloud infrastructure costs through 2029.
This translates to a baseline average of $16 bn annually dedicated solely to leasing the silicon necessary to host and serve the Claude models to the public.
Furthermore, OpenAI’s long-term capital expenditure plans indicate a staggering $1.4 trillion commitment to datacenter infrastructure projects over the next eight years.
While these are capital investments, the depreciation of these assets forms a massive, unavoidable line item on the income statement that must ultimately be recovered through user subscription revenue.
Pillar 2: Model Training and Amortized Capital Expenditures
Inference - the computational cost of actively generating responses for users in real-time - is only half of the compute equation.
Historical allegations indicate that the GPT-4, released in 2023, cost approximately $100 million to train utilizing a cluster of 10,000 Nvidia A100 processors.
GPT-5 class model is estimated to cost upwards of $500 million in raw compute time alone, excluding the salaries of the researchers and the cost of the underlying data.
For the calendar year 2026, OpenAI plans to spend an estimated $32 bn specifically on model training, a figure that is projected to scale to $65 bn by 2027.
Under a traditional venture-subsidized model, this immense capital expenditure is floated by continuous fundraising rounds.
However, in an unsubsidized, free-market unit economic model, this continuous capital injection does not exist. The $32 bn spent on training in 2026 cannot be written off as speculative R&D. It must be treated as a capitalized software development expense and amortized across the paying user base over the useful life of the model.
Given that the half-life of a state-of-the-art AI model is currently <12m before it is rendered obsolete by a successor, this schedule is incredibly aggressive, placing an immense burden on the monthly subscription fee.
Pillar 3: The Talent War and Human Capital Liabilities
To acquire and retain this talent against the gravitational pull of hyperscalers like Meta, Google, and Apple, compensation packages at frontier AI laboratories have skyrocketed, driven almost entirely by equity grants.
In 2025, OpenAI’s workforce of approximately 4,000 to 4,500 employees received stock-based compensation averaging $1.5 million per worker.
This average is heavily skewed by multi-million dollar retention grants, accelerated vesting schedules, and aggressive one-time payouts for critical leadership roles.
To contextualize, OpenAI's average stock compensation per employee is roughly 34x the average stock compensation of 18 large technology companies in the year prior to their respective initial public offerings.
In the first half of 2025 alone, OpenAI recognized nearly $2.5 bn in stock compensation, alongside $6.7 bn in broader R&D spending.
Extrapolated annually, this represents an approximate $5 bn to $6.75 bn human capital expense for OpenAI in 2026.
Anthropic operates with a slightly leaner, though rapidly expanding, workforce.
The company grew from just 192 employees in 2022 to over 1,097 by 2025, representing a 471% workforce expansion, and is projected to house approximately 2,500 employees by 2026.
The average total compensation at Anthropic is intensely competitive, ranging from $495,000 to $560,000 per year, heavily weighted toward equity and share-based payments.
For an organization of 2,500 employees, this represents a baseline annual human capital expenditure of approximately $1.25 bn to $1.4 billion.
While equity compensation does not drain liquid cash reserves immediately, in an unsubsidized, publicly accountable unit economic model, the dilution and eventual liquidity requirements represent a real, systemic cost.
Pillar 4: SG&A, Revenue Sharing, and Intellectual Property Licensing
OpenAI has proactively partnered with over 20 global news organizations, including The Washington Post, News Corp, Time, and Axel Springer, to license high-quality training data and integrate real-time, copyrighted search capabilities into its models.
While smaller deals reportedly range from $1 million to $5 million annually, marquee partnerships cost significantly more.
These agreements establish a permanent, recurring COGS for high-quality data ingestion that did not exist in the GPT-3 era.
Anthropic faces similar legal and compliance overhead, underscored by high-profile copyright infringement lawsuits filed by entities like music publishers.
Structural corporate agreements act as a massive drag on gross margins.
From an agreement penned in October 2025, OpenAI is obligated to pay Microsoft 20% of its total revenue through the year 2032.
In 2024, Microsoft received $493.8 million in revenue share payments.
This figure climbed to $865.8 million in just the first three quarters of 2025.
For a projected 2026 revenue of $30 bn , this 20% revenue share represents a staggering $6 bn liability. This functions as an unavoidable structural tax on OpenAI’s gross revenue, dramatically increasing the breakeven threshold for every individual subscriber.
OpenAI: The 2026 Unsubsidized Break-Even Subscription Price
Establishing the Operational Cost Baseline for 2026
Based on the aggregated financial disclosures and leaks, OpenAI’s projected expense baseline for the calendar year 2026 can be modeled as follows:
User Base and Conversion Dynamics
OpenAI’s top-of-funnel user base has experienced staggering growth.
~900 million weekly active users (WAUs) by early 2026,
~50 million of those being paying subscribers.
Under the current subsidized model, the massive cost of serving the remaining 850 million free users is floated by venture capital, allowing OpenAI to maintain market dominance and continuous pipeline of human reinforcement learning data. However, under the strict premise of an unsubsidized environment, the core fixed costs of the company must be borne entirely by the active, paying subscriber base.
The Unsubsidized OpenAI Cost Per User Calculation (Baseline)
Calculate Total Fixed Costs:
Training ($32.0B) + Human Capital ($13.4B) + Microsoft Tax ($6.0B) + SG&A ($2.0B) = $53.4 bn in Fixed Costs.Calculate Fixed Cost Per User:
$53.4 bn / 50 Million Users = $1,068.00 per user, per year.
Divided by 12 months = $89.00 per month in fixed cost burden per user.Calculate Variable Inference Cost Per User: The $9.3 bn in premium inference costs divided by 50 million users equals $186.00 per year, or $15.50 per month on average.
Assuming a blended variable inference cost of $50 per month for a highly active, unsubsidized premium user (factoring in advanced reasoning models), the baseline is:
Fixed Cost Burden: $89.00 / month
Variable Inference Cost: $50.00 / month
Baseline Break-Even Subscription Price: $139.00 per month per account.
Factoring in Debt and Long-Term Infrastructure Obligations
While the $139.00 figure covers core operational baseline costs, it ignores the immense weight of long-term debt and infrastructural lock-ins.
OpenAI has committed an estimated $1.4 trillion to data center infrastructure over the next eight years which amortizes to $175 bn annually.
The company recently secured a $4 bn revolving credit facility at SOFR + 100 basis points (~6% as April 2026). This adds $240 million in annual debt servicing.
Adding this $175.24 bn burden to the $62.7 bn operational baseline pushes the total annual revenue requirement to roughly $238 billion.
Divided across the 50 million paying subscribers, the fixed cost equates to $4,760 per user, per year, or $396.66 per month.
When combined with the $50 blended variable inference cost, the true unsubsidized break-even price soars to $446.66 per month.
Anthropic: The 2026 Unsubsidized Break-Even Subscription Price
By late 2025, Anthropic had >300k business customers, which accounted for approximately 80% of its total revenue.
The company saw large accounts >$100k annually grow sevenfold in a year, with >500 customers spending in >$1 million annually.
This B2B strategy culminated in an explosive revenue curve, from $10 million in 2022 to $1 bn ARR in late 2024, to $4 bn by mid-2025, and accelerating to an annualized revenue run rate of $19 bn by March 2026.
A primary driver of this growth is Claude Code, an agentic coding assistant that reached $2.5 bn in ARR by early 2026.
Establishing the Operational Cost Baseline for 2026
The Unsubsidized Anthropic Cost Per User Calculation (Baseline)
Estimates place Anthropic's total consumer user base between 18-30 million in 2026. Because the vast majority of Anthropic's revenue is derived from enterprise and API customers, we can assume that roughly 15 million of these users are monetized seats, power developers, or API clients.
Calculate Fixed Cost Per User:
$23.3 bn / 15 Million Users = $1,553.33 per user, per year.
Divided by 12 months = $129.44 per month in fixed cost burden per user.Calculate Variable Inference Cost Per User: Unlike consumer chatbots, Anthropic's ecosystem is dominated by software developers. Code generation is notoriously token-intensive, consuming 10 to 50 times more tokens than typical chat interactions.
If an average enterprise developer utilizes 50 million cache read tokens ($15.00), 10 million standard input tokens ($30.00), and 5 million output tokens ($75.00) in a month using Claude 3.5 Sonnet, their variable inference cost is $120.00 per month.Total Baseline Break-Even Calculation:
Fixed Cost Burden: $129.44 / month
Variable Inference Cost: $120.00 / month
Baseline Break-Even Subscription Price: $249.44 per month per account.
Factoring in Custom Infrastructure and Revenue-Sharing Obligations
$50 bn investment in custom data centers in partnership with Fluidstack. Amortized over a standard four-year hardware cycle, this adds $12.5 bn annually.
Aggressive revenue-sharing agreements with its cloud partners, paying ~$1.9 bn in 2026 (scaling to $6.4 bn by 2027) back to providers like Amazon and Google.
Adding the $12.5 bn infrastructure amortization and the $1.9 bn cloud revenue share to their $23.3 bn baseline pushes Anthropic's annual operational requirement to roughly $37.7 billion.
Divided across 15 million paying users, the fixed cost equates to $2,513 per year, or $209.41 per month.
Combined with the $120 variable inference cost, the unsubsidized baseline price jumps to $329.41 per month for standard professional access.
Heavy enterprise API users, running extensive iterative coding sessions, would still easily scale into the thousands of dollars.
Projecting the 2029 Horizon: The Trillion-Dollar Squeeze
As AI platforms scale toward the end of the decade, the infrastructural demands and financial stakes multiply, making the path to profitability even more arduous. Projections for 2029 highlight a widening chasm between anticipated revenue and the compounding costs of AI development and compute.
According to Bloomberg and internal forecasts, OpenAI is targeting $125 bn in annual revenue to reach break-even profitability by 2029 (Lipshultz, 2025).
However, this revenue target is counterbalanced by catastrophic spending projections.
OpenAI 2029: Calculating the Unsubsidized Break-Even Subscription Price
According to Bloomberg and internal forecasts, OpenAI is targeting $125 bn in annual revenue to reach break-even profitability by 2029.
Note: These are internal projections which will be optimistic
If we project OpenAI successfully scaling its paid user base to an estimated 200 million accounts in 2029 (on track for 220 million by 2030), the unsubsidized calculation would be:
Calculate Fixed Cost Per User:
$300.0 bn / 200 Million Users = $1,500.00 per user, per year.
Divided by 12 months = $125.00 per month in fixed cost burden per user.Calculate Variable Inference Cost Per User:
As models advance toward AGI capabilities (e.g., GPT-6 or GPT-7), despite hardware becoming cheaper per token, the volume of tokens burned per session will surge. A conservative estimate for a heavy power user running continuous autonomous agents by 2029 places inference at $100 per month.Total Unsubsidized 2029 Break-Even Calculation:
Fixed Cost Burden: $125.00 / month
Variable Inference Cost: $100.00 / month
Unsubsidized Break-Even Subscription Price: $225.00 per month per account.
When factoring in the exponentially heavier variable inference costs associated with executing future reasoning models (such as GPT-6 or GPT-7) and continuous data center amortizations, the unsubsidized price floor for a power user in 2029 realistically remains between $250 and $400 per month.
Anthropic 2029: Calculating the Unsubsidized Break-Even Subscription Price
A significant portion of this cost is driven by their $80 bn multi-year cloud compute commitment to partners like Amazon and Google running through 2029. Furthermore, their revenue-sharing agreements are expected to cost $6.4 bn by 2027 alone, and will likely scale to over $10 bn annually by 2029.
If Anthropic's enterprise-heavy model successfully scales to an estimated 30 million monetized enterprise and developer seats by 2029, the unsubsidized calculation would be:
Calculate Fixed Cost Per User:
$93.5 bn / 30 Million Users = $3,116.66 per user, per year.
Divided by 12 months = $259.72 per month in fixed cost burden per user.Calculate Variable Inference Cost Per User:
For a platform geared heavily toward autonomous software engineering, the tokens consumed by background "coding agents" managing entire codebases 24/7 will be immense. The variable inference compute requirement is estimated at $150.00 per month.Total Unsubsidized 2029 Break-Even Calculation:
Fixed Cost Burden: $259.72 / month
Variable Inference Cost: $150.00 / month
Unsubsidized Break-Even Subscription Price: $409.72 per month per account.
These 2029 projections confirm that the AI unit economic problem does not soften with scale; it steepens with adoption.
The 2-Year Hardware Cycle: Recalculating Break-Even Prices for 2026 and 2029
To understand the true, unsubsidized cost of delivering artificial intelligence, we must recalculate the unit economics under a strict two-year hardware refresh cycle. If the six-year accounting illusion is stripped away, the financial load carried by each paying subscriber surges drastically.
The 60/40 Infrastructure Split
A modern AI data center is not entirely composed of fast-depreciating silicon. According to industry capital expenditure breakdowns, approximately 60% of data center investments go toward specialized technology and hardware (chips, accelerators, servers), while the remaining 40% goes toward longer-lived assets like land, site development, mechanical cooling, and power generation. By separating the rapidly obsolescing GPU modules from the physical data center shells, we can apply a strict two-year depreciation schedule to the 60% hardware component, while maintaining the longer multi-year amortizations for the 40% facility component.
OpenAI: 2026 and 2029 Strict 2-Year Amortization
OpenAI's estimated $1.4 trillion infrastructure commitment can be split into $840 bn for hardware (60%) and $560 bn for facilities (40%).
If that $840 bn in silicon must be refreshed and amortized every two years, the hardware burden alone becomes $420 bn annually.
The facilities, amortized over the original eight-year timeline, add $70 bn annually, creating a new total infrastructure burden of $490 bn per year (a massive leap from the previously blended $175 bn annual average).
2026 Recalculation: Adding this $490 bn infrastructure load to OpenAI's $62.7 bn operational baseline and $0.24 bn debt service pushes the 2026 annual revenue requirement to roughly $553 billion.
Divided across 50 million paying users, the fixed cost equates to $11,060 per year, or $921.66 per month. Adding the $50 in variable inference results in a true unsubsidized break-even price of $971.66 per month for 2026.
2029 Recalculation: Assuming OpenAI hits its $125 bn operational target and scales to 200 million paid users by 2029. Adding the $490 bn hardware and facilities load pushes the total annual requirement to $615 billion.
Across 200 million users, the fixed cost is $3,075 annually, or $256.25 per month. Combined with $100 in future variable inference costs, the 2029 unsubsidized break-even price settles at $356.25 per month.
Anthropic: 2026 and 2029 Strict 2-Year Amortization
Anthropic's $50 bn custom data center buildout with Fluidstack breaks down to $30 bn in hardware and $20 bn in facilities. Amortizing the hardware over two years ($15 bn annually) and the facilities over four years ($5 bn annually) raises their custom infrastructure baseline from $12.5 bn to $20 bn annually.
2026 Recalculation: The baseline operational cost ($23.3 billion) plus the revised custom infrastructure ($20 billion) and cloud revenue share ($1.9 billion) totals $45.2 billion.
Distributed across 15 million paying users, the fixed cost is $3,013.33 annually, or $251.11 per month. With a $120 variable inference cost, the unsubsidized 2026 break-even price reaches $371.11 per month.
2029 Recalculation: Taking the $71 bn operational baseline for 2029, adding the $20 bn 2-year infrastructure cycle and a projected $10 bn cloud revenue share brings the total requirement to $101 billion.
Across 30 million users, the fixed cost is $3,366.66 annually, or $280.55 per month. Combined with $150 in variable inference, the 2029 break-even price pushes to $430.55 per month.
The Impact of Agentic Workflows on Variable Inference Costs
The economic disparity between the current $20 subscription fees and the $300+ unsubsidized realities is primarily driven by the architectural shift toward agentic workflows.
GenAI is no longer a passive retrieval mechanism; it is an active computational engine.
The market has seen API inference costs decline 10x annually, with GPT-4 equivalent performance now costing $0.40 per million tokens versus $20 in late 2022.
However, Jevons Paradox dictates that as technological efficiency increases, the demand for the resource increases proportionately.
While the cost per token has plummeted, the volume of tokens demanded by the models has skyrocketed. Models like Claude Opus 4.6 and GPT-5 feature massive context windows (up to 1 million and 400,000 tokens, respectively).
When a user asks an agent to review a pull request, the model must ingest the entire codebase, generate thousands of hidden "thinking" tokens to reason through the logic, execute tool calls to run tests, and finally output the synthesized code.
Because the entire context must be re-fed into the system for each sequential query , a 50-message thread utilizes five times more processing power than five 10-message chats. This exponential token burn means that even with 10x cheaper hardware, the variable inference cost per user session is increasing.
The economics of "cognitive offloading" dictate that as users delegate more complex, multi-step tasks to AI agents, the computational load—and thereby the unsubsidized cost to serve that user—will continue to rise.
Stress-Testing the 2029 Unit Economics: Four Risk Scenarios
Scenario 1: Interest Rates Double
If the 10-year Treasury yield doubles from 4.345% to 8.69%, the environment for AI funding changes. The AI boom is heavily reliant on debt markets and syndicated loans.
OpenAI's multi-bn dollar revolving credit facilities are tied to the SOFR plus a margin; doubled rates would instantly multiply direct debt servicing costs
An increased risk-free rate alters the equity risk premium. Under this scenario, we assume the VC markets contract and half of the projected $100B+ external infrastructure funding dries up.
OpenAI: annual capital and debt servicing shortfall of $25 bn (due to dried-up VC markets and higher debt costs), this adds to $615 bn burden. The $640 bn across 200 million users adds $10.41/month. The breakeven price is $366.66 /month.
Anthropic: A proportionate $10 bn annual capital and debt shortfall raises their burden to $111 billion. Across 30 million users, this adds $27.77 per month. The new break-even price jumps to $458.32 per month.
Scenario 2: Utility (Electricity and Water) Costs Increase by 50%
Energy is the fundamental constraint of AI scale. Operating a 1-gigawatt (GW) data center consumes over $1.3 bn in electricity annually at standard rates. Both companies have massive power pipelines: OpenAI targets at least 10 GW of capacity, while Anthropic's expanded cloud partnerships bring well over 1 GW of capacity online. If utility rates spike by 50%—driven by grid strain and the massive energy draw of AI facilities—the operational expenditure balloons:
OpenAI (10 GW): Baseline electricity costs of roughly $13 bn jump to $19.5 billion. This $6.5 bn increase divided by 200 million users adds $2.70 per month, raising the break-even to $358.95 per month.
Anthropic (1.5 GW estimated): Baseline electricity costs of roughly $1.95 bn jump to $2.92 billion. The nearly $1 bn increase divided by 30 million users adds $2.77 per month, raising the break-even to $433.32 per month.
Scenario 3: Infrastructure Costs Increase by 50%
Supply chain bottlenecks, global GPU scarcity, and raw material shortages could easily drive the cost of building and equipping AI data centers higher. If the cost of the underlying hardware and facilities increases by 50% under the strict 2-year amortization cycle, the unit economics suffer catastrophically:
OpenAI: A 50% premium on their $490 bn annual infrastructure burden adds $245 bn in yearly fixed costs. Spread across 200 million users, this adds an immense $102.08 per month. The new break-even price surges to $458.33 per month.
Anthropic: A 50% premium on their $20 bn annual custom infrastructure burden adds $10 bn in yearly fixed costs. Spread across 30 million users, this adds $27.77 per month. The new break-even price jumps to $458.32 per month.
Scenario 4: Projected 2029 User Base Falls Short by 25%
The most sensitive variable in the unit economics model is user volume. If AI adoption plateaus or churn rates increase due to pricing fatigue, the massive fixed costs must be borne by a smaller cohort of subscribers:
OpenAI: If the projected 200 million paid users fall short by 25% to 150 million, the $615 bn in total fixed costs becomes much heavier per capita. The fixed cost burden rises from $125.00 to $341.66 per month. When combining the $100 variable inference cost, the new break-even price spikes by $85.41, reaching $441.66 per month.
Anthropic: If the projected 30 million enterprise users fall short by 25% to 22.5 million, their $101 bn in total fixed costs becomes heavily concentrated. The fixed cost burden rises from $280.55 to $374.07 per month. Adding the $150 variable inference cost pushes the new break-even price up by $93.52, hitting $524.07 per month.
AI Tokens and Pricing
Flat-rate AI subscription is economically unsustainable because user consumption is outpacing model efficiency gains.
While the cost per token is falling, the cost per user is rising (Ding, 2025).
(Ding, 2025).
While providers often tout that new models are 10x cheaper per token than previous versions (Koparkar & Koparkar, 2026), this doesn't result in lower monthly costs.
As models become cheaper and faster, developers build more complex workflows.
Reasoning Models Burn Massive Compute: The industry shift from standard LLMs to "reasoning" models (such as OpenAI's o-series, DeepThink, or Grok's reasoning mode) fundamentally changes token consumption. These models generate massive amounts of "hidden" output tokens to "think" through a problem before providing the final answer to the user. (Theo - t3․gg, 2025)
These workflows use more tokens (1000x +) to complete a single task compared to a simple chat prompt, effectively removing the efficiency gains
Theo's experiments and the referenced research highlight that modern models are structurally designed to burn significantly more tokens for the same prompts.
In one highlighted example, a reasoning model burned over 600 tokens just to generate a two-word response. As AI is given more complex tasks, token usage is scaling exponentially. (Theo - t3․gg, 2025)
Users naturally gravitate toward the most capable "frontier" models. When a user switches from an older, cheaper model to a newer, more expensive one, the provider's margin shrinks or disappears instantly.
The Death of Zero Marginal Cost: Unlike traditional SaaS software where adding one more free user costs practically nothing, every single AI interaction carries a hard compute cost. (Theo - t3․gg, 2025)
AI companies are caught between two pressures:
Market Expectation
Consumers have been trained by Netflix and Spotify to expect unlimited flat-rate pricing.
Rising Input Costs
Power users and complex AI agents can easily consume $1000s worth of compute in a single month.
As a result, companies offering unlimited tiers are "short" on tokens and "long" on fixed-price subscriptions, leading to massive losses on their most active users.
Recent industry shifts as proof of this squeeze (Ding, 2025).
Popular AI coding tools have had to adjust pricing or implement limits because power users were burning through $1000s in compute.
Cursor
Windsurf
Companies are afraid to be the first to switch to pure usage-based pricing for fear of losing users to VC-subsidized competitors
To survive, companies will likely move toward usage-based pricing, hybrid models (where some tasks are processed locally on the user's device), or outcome-based pricing (charging for the task completed rather than the tokens used) (Koetsier, 2025)(Morris, 2026)(Yeung, 2026)(Anderson, 2026)
Hypergrowers Exposure
NVIDIA
Revenue Already "Captured" (Deferred Revenue)
Total Deferred Revenue: $2,572 million.
Customer Advances: $160 million of the above total represents cash already in-hand.
Additions in FY2026: During the fiscal year, NVIDIA added $11,137 million to its deferred revenue balance, which included $9.0 billion in new customer advances.
Potential Exposure (Purchase and Lease Obligations)
The primary risk to NVIDIA in the event of massive cancellations is the scale of its non-cancelable commitments to its own suppliers and partners:
Inventory Purchase & Supply Obligations: $95.2 billion, substantially all of which is due to be paid through fiscal year 2027.
Multi-year Cloud Service Commitments: $27.0 billion through 2032 to support research and development.
Facility Lease Guarantees: NVIDIA has guaranteed up to $3.5 billion of its partners' facility lease obligations.
Liquidity Buffer
NVIDIA maintains a substantial cash position to weather potential market volatility or contract cancellations:
Cash, Cash Equivalents, and Marketable Securities: $62.6 billion as of January 25, 2026.
Risk Summary
NVIDIA is "at risk" if cancellations occur at a scale where its $62.6 billion in liquidity plus the $160 million in existing customer advances cannot cover its $95.2 billion in non-cancelable supply obligations. While a large portion of its forward revenue is based on purchase orders that customers can generally cancel or change with little notice, the company uses its massive cash reserves and customer prepayments to mitigate the impact of its own non-cancelable obligations to foundries like TSMC.
(NVIDIA, 2026)
MSFT & AMZN
Microsoft
For 2026, its global AI infrastructure CapEx is projected to reach $100-$120 billion.
Microsoft secured a $14 billion deal with Nscale to supply an additional 116,000 NVIDIA GB300 GPUs across the U.S. and Europe.
In April 2026, Microsoft entered into a $9.7 billion agreement with data center operator IREN to secure further access to NVIDIA's GB300 AI chips, including a 20% prepayment to lock in supply.
CFO Amy Hood noted the company remains "capacity-constrained through at least the end of the fiscal year,"
Strategy of doubling its data center footprint and adding 80% more AI capacity in 2026.
Amazon
In February 2026, Amazon CEO Andy Jassy shocked Wall Street by announcing a record $200 billion capital expenditure plan for the year, with the "vast majority" dedicated to AWS infrastructure.
While Amazon is aggressively building its own custom chips (Trainium and Graviton), it maintains a 15-year partnership with NVIDIA. Analysts expect NVIDIA to be a primary beneficiary of the $500 million per day Amazon is spending on infrastructure.
Jassy noted that nearly all of its Trainium 3 supply is already committed through mid-2026, forcing a continued reliance on NVIDIA to fill the gap.
Anthropic (Project Rainier) involves a massive deployment of at least 500,000 Trainium chips, which ironically increases its need for NVIDIA-based networking and software integration to manage the hybrid fleet.
Comparative Risk Profiles: The "CapEx vs. Free Cash Flow" Gap
The aggressive spending has fundamentally changed the risk profiles of these two giants:
Because these companies have essentially become the bankers, landlords, and sole suppliers to the AI startup ecosystem, they are exposed to a "circular" contagion.
Financial Risk Profiles
Microsoft
45% of its $625 billion cloud backlog is tied to OpenAI.
If OpenAI fails to reach profitability or faces obsolescence, nearly half of Microsoft’s projected future revenue could vanish from its books overnight.
The SEC is currently scrutinizing whether Microsoft’s revenue from OpenAI is real. Because Microsoft invests cash that OpenAI then immediately "spends" back on Azure, any regulatory reclassification of these "compute credits" would force a massive restatement of Microsoft’s cloud growth.
Microsoft has billions in paper gains on its OpenAI investment. If OpenAI’s valuation (currently targeting $850B+) collapses in a public market, Microsoft would have to take a multi-billion dollar non-cash impairment charge.
Amazon
$500 million per day on data centers and custom chips (Trainium). If AI demand cooling occurs, Amazon will be left with specialized, power-hungry buildings and chips that cannot be easily repurposed
Amazon could see negative free cash flow of $17B to $28B in 2026 due to this spending. Unlike Microsoft, which has a software "cash cow," Amazon’s retail margins are thin (under 6%)
$50 billion deal with OpenAI and $8 billion with Anthropic makes it a major creditor. If these startups enter insolvency Amazon loses not just an investment, but its largest "anchor tenants" for its new data centers.
Risk Scenarios: What Happens if Things Go Sideways?
Potential Strategic Outcomes
The "Hard Takeover": If OpenAI or Anthropic faces bankruptcy, Microsoft and Amazon will not let them die. They will likely use their "Senior Creditor" status to forcibly acquire the intellectual property and teams for pennies on the dollar, effectively turning the startups into internal "research divisions" (similar to how Microsoft absorbed the Inflection AI team in 2024).
To appease shareholders during slowing growth, they may be forced to initiate (or significantly hike) dividends and slash R&D spending in other areas
To save margins, they will stop buying NVIDIA chips entirely and force all customers onto their own, cheaper "proprietary" chips
Requirements
1. Hardware & Compute Infrastructure (Highest Cost)
Cost Scale: Tens of Billions of Dollars (CapEx) The absolute largest expense for AI companies is the physical hardware required to train and run models. This primarily consists of massive clusters of AI accelerators (like NVIDIA H100s/B200s or Google TPUs), networking infrastructure (InfiniBand/Ethernet), and the physical data centers to house them.
Detailed Breakdown of Semiconductors & Infrastructure:
Silicon Wafers: The foundation of chips. High-purity quartz sand is mined (often from places like Spruce Pine, North Carolina, which produces the world's purest quartz), refined into polysilicon, and grown into ingots. These are sliced into wafers primarily by companies in Japan (Shin-Etsu) and Taiwan (GlobalWafers).
High-Bandwidth Memory (HBM): AI chips require memory stacked directly on the processor to feed data fast enough. HBM is almost exclusively manufactured in South Korea by SK Hynix and Samsung.
Trace Metals & Rare Earths:
Copper & Gold: Used for microscopic interconnects and pins. Sourced globally (Chile, Peru, Australia).
Tantalum: Used for high-end capacitors. Often sourced from Central Africa and Australia.
Gallium & Germanium: Used in advanced power electronics and optoelectronics. Currently dominated by China, posing a geopolitical supply chain risk.
Manufacturing Apparatus (Lithography): The chips cannot be printed without Extreme Ultraviolet (EUV) lithography machines made solely by ASML in the Netherlands. These machines rely on ultra-precise mirrors made by Zeiss in Germany, which in turn use specialized lasers from Trumpf (also Germany).
Fabrication: The actual printing of the world's most advanced AI chips is almost entirely concentrated at TSMC foundries in Taiwan.
Networking Gear: Millions of miles of fiber optic cables (glass, Kevlar, polyethylene) and advanced optical transceivers are needed to link tens of thousands of GPUs together so they can act as a single supercomputer.
2. Energy & Cooling (Power Generation)
Cost Scale: Hundreds of Millions to Billions of Dollars (OpEx) Once the multi-billion-dollar supercomputers are built, they require staggering amounts of power. Training a single frontier model requires running up to 100,000 GPUs at maximum capacity for months. Furthermore, inference (answering user queries) is beginning to cost more than training as user bases scale into the hundreds of millions.
Detailed Breakdown of Power Needs:
Electricity (Baseload Power): AI data centers cannot rely solely on intermittent renewables (solar/wind) because they must run 24/7. They require massive baseload power. This is currently fulfilled by natural gas (sourced via domestic pipelines in the US) and coal. Increasingly, AI companies are investing heavily in Nuclear Power (e.g., Microsoft’s deal to restart Three Mile Island) to secure zero-carbon baseload power, which requires Uranium (sourced from Kazakhstan, Canada, and Australia).
Cooling Systems (Water & HVAC): Chips running at high utilization melt without intense cooling.
Water: Data centers consume millions of gallons of fresh water for evaporative cooling, sourced from local municipal watersheds.
Metals: Liquid cooling systems require massive amounts of copper piping, aluminum heat sinks, and specialized synthetic refrigerants (chemicals like fluorocarbons produced by industrial giants like 3M or Chemours).
3. Precision Manufacturing Inputs (The "Non-Negotiables")
Cost Scale: Billions of Dollars per Fabrication Plant (Fab)
Ultra-High Purity (UHP) Gases: UHP Hydrogen (6N or 7N purity) is strictly required to react with tin debris and keep the EUV mirrors clean. However, fabs also require massive, continuous pipelines of UHP Nitrogen, Argon, and Helium to create perfectly inert environments so silicon doesn't oxidize during the hundreds of baking and etching steps.
The Sourcing: These gases cannot be easily shipped in their highest purity states over long distances without risking contamination, so they are often generated near or directly on the fab site by massive industrial gas conglomerates.
Key Players: The market is dominated by a tight oligopoly: Linde (UK/Germany), Air Liquide (France), Air Products (USA), and Taiyo Nippon Sanso (Japan).
The Helium Exception: Helium, crucial for cooling and inert environments, cannot be synthesized; it must be extracted from the earth alongside natural gas. The vast majority of the world's supply comes from just a few countries: the United States (historically the Texas/Oklahoma panhandle), Qatar, and Algeria.
The World's Smoothest Optics: The EUV light is created at a wavelength of 13.5 nanometers. This light is absorbed by almost everything, including air and standard glass lenses. Therefore, the light must be bounced off a series of Bragg reflectors (specialized mirrors) in a vacuum.
The Requirement: These mirrors, manufactured by Zeiss in Germany, must be so flawlessly smooth that if one were blown up to the size of the continental United States, the tallest "mountain" (imperfection) on its surface would be roughly the size of a millimeter.
The Sourcing: There is essentially only one company on earth capable of manufacturing the Bragg reflector mirrors to the atomic precision required for ASML’s EUV machines.
Key Player: Carl Zeiss SMT, located in Oberkochen, Germany. Their mastery of precision optics is a primary bottleneck; ASML cannot build more EUV machines than Zeiss can build mirrors.
Extreme Lasers and Pure Tin: To actually generate the EUV light, a high-power CO2 laser (built by Trumpf) is fired at microscopic droplets of molten tin.
The Requirement: The machine must perfectly shoot a falling tin droplet with a laser to flatten it, and then shoot it again a microsecond later to vaporize it into plasma. It does this 50,000 times per second.
The Lasers: The high-power $CO_2$ lasers used to blast the tin are exclusively engineered by TRUMPF, located in Ditzingen, Germany.
The Droplet Generator: The system that perfectly drops the molten tin 50,000 times a second was developed by Cymer, a company based in San Diego, California (which ASML bought outright to secure the technology).
The Tin: While raw tin is mined globally (heavily in China, Indonesia, and Peru), the ultra-pure, electronics-grade tin required for the EUV plasma process is refined by specialized chemical companies in Japan and the United States.
Ultra-Pure Water (UPW): After wafers are etched or polished, they must be washed. Normal tap water or even standard filtered water contains microscopic minerals, bacteria, and ions that would short-circuit a microscopic transistor.
The Requirement: Fabs require highly processed UPW. It is essentially pure H2O with all ions removed. It is so pure that it acts as a hungry solvent—if you drank it, it would actively leach essential minerals out of your teeth and bones. Fabs consume millions of gallons of this a day.
The Sourcing: UPW degrades instantly if exposed to air or standard pipes, so it cannot be shipped. It must be manufactured on-site at the fab. Fabs are strategically built near robust municipal water sources (rivers, lakes, or large reservoirs) because they consume millions of gallons daily.
The Processing Systems: While the water is local, the massive, multi-stage filtration systems (reverse osmosis, UV oxidization, ion exchange) required to turn tap water into UPW are built by specialized companies like Kurita Water Industries (Japan), Evoqua/Xylem (USA), and Ovivo (Canada).
Absolute Seismic and Vibration Isolation: Because the machines are printing features that are only a few atoms wide, even the slightest vibration will ruin a $30,000 silicon wafer.
The Requirement: EUV machines must be suspended on active dampening systems. The fab floors are often isolated from the rest of the building. They are so sensitive that loud noises, a truck driving by outside, or even tiny micro-seismic tremors from miles away can disrupt the yield if not properly isolated.
The Sourcing: The active dampening systems that keep the machines perfectly still are highly specialized pieces of structural engineering.
Key Players: The components (pneumatic isolators, active piezoelectric cancellation systems) are sourced from precision engineering firms like TMC / Ametek (USA), Kinetic Systems (USA), and Bilz (Germany).
Class 1 Cleanrooms: A single speck of dust landing on a wafer during printing is like a boulder crushing a city grid at the nanoscale.
The Requirement: The manufacturing floor must be a "Class 1" or better cleanroom, meaning there is roughly less than one particle of dust (larger than 0.5 microns) per cubic foot of air. For context, normal room air has millions of particles per cubic foot. This requires immense HVAC, HEPA filtration, and strict protocols where humans are completely sealed in "bunny suits" to prevent skin cells or hair from escaping.
The Construction: Building a Class 1 cleanroom is a massive architectural feat. The leading company globally for designing and building these ultra-clean fab environments is Exyte, based in Stuttgart, Germany.
The Filtration: The vital HEPA and ULPA filters that constantly scrub the air are sourced from companies like Camfil (Sweden) and American Air Filter (AAF), a subsidiary of Daikin (Japan).
The "Bunny Suits": The specialized, lint-free, static-dissipative garments worn by workers to protect the silicon from human contamination are manufactured by chemical and materials giants like DuPont (USA) and Kimberly-Clark (USA).
Risks
The global AI boom is currently colliding with the hard physical limits of the real world. Based on the current landscape in 2026, the supply chain has shifted from being constrained by capital to being constrained by physical infrastructure and geopolitics.
Here are the biggest, most critical risks threatening each layer of the AI supply chain today:
1. Hardware & Compute Infrastructure Risks
The High-Bandwidth Memory (HBM) Drought: AI's insatiable demand for memory has fundamentally outpaced global manufacturing capacity. Companies are facing severe HBM shortages that are drastically extending lead times for AI servers (shifting from typical 8-week waits to 6 months or more) and causing massive price inflation across the board.
Foundry Capacity Ceilings: Production remains dangerously concentrated. Top foundries like TSMC are hitting hard capacity limits for advanced AI chips. Because building new fabrication plants takes years, the industry is stuck in a "zero-sum" bottleneck where AI hardware is cannibalizing the manufacturing capacity needed for other technology sectors.
Geopolitical Trade Wars & "Shadow Markets": Weaponized export controls (like US restrictions on advanced tech to China) are creating immense friction. This has spawned a multi-billion-dollar black market of smuggled GPUs and servers, adding deep unpredictability and legal risk to the global hardware ecosystem.
2. Energy & Cooling Risks
Grid Congestion & Energization Delays: Power is now the absolute primary constraint on AI scaling. New gigawatt-scale data centers are facing multi-year delays simply waiting in interconnection queues to be attached to severely strained municipal power grids.
The "Baseload" Dilemma: AI requires 24/7 uninterrupted power, which intermittent renewables (wind/solar) cannot independently provide. The scramble for firm baseload power is forcing companies to increasingly rely on natural gas or highly complex, long-term nuclear projects, creating tension between the need to scale and corporate climate pledges.
Regulatory Backlash & "Kill Switches": As data centers begin to consume electricity at the scale of entire nations, local governments are pushing back. Emerging regulations—such as mandates to match demand with local generation or emergency grid "kill switches" that can cut data center power to protect civilian infrastructure—directly threaten operational uptime.
3. Precision Manufacturing Inputs Risks
Critical Material Chokepoints (The Tungsten & Rare Earths Threat): Advanced chipmaking relies heavily on materials dominated by single nations. For example, recent export restrictions by China on tungsten—a dense metal essential for advanced node manufacturing with no practical substitute—have caused historic price rallies and severe forecasted supply deficits.
"Single Point of Failure" Equipment: The entire ecosystem relies on a handful of hyper-specialized companies (like Zeiss for atomic-level optics or ASML for lithography). An energy crisis, cyberattack, or natural disaster at just one of these facilities can stall global chip production entirely, as there are zero alternative suppliers.
Transit & Chokepoint Vulnerabilities: Foundational inputs like ultra-high purity gases (e.g., helium) and specialty chemicals must navigate fragile global shipping lanes. Regional conflicts and disruptions to major maritime chokepoints (like the Strait of Hormuz) can instantly cut off the raw materials needed to keep fabrication plants online.
Nvidia (NVDA)
Amount of Exposure: As of early 2026, Nvidia's data center segment accounts for roughly 90% of its revenue base (reporting around $215.9 billion for FY2026). Its financial growth is almost entirely tethered to AI infrastructure build-outs.
Critical Risk Scenario: Nvidia's premium valuation heavily prices in sustained 40%+ revenue growth. If hyperscalers moderate their spending or delay next-generation chip orders (like the Blackwell or Rubin architectures) by just 15% to 20%—translating to roughly a $30 billion to $40 billion revenue shortfall—it could trigger a severe multiple contraction. A drop to 20% annual growth would deeply impact its market capitalization.
The Hyperscalers (Alphabet, Microsoft, Meta, Amazon)
Amount of Exposure: These tech giants are exposed via historic capital expenditures. Collectively, they are projected to spend between $500 billion and $660 billion on AI infrastructure in 2026 alone. Alphabet, for instance, targeted up to $185 billion in CapEx for the year.
Critical Risk Scenario: The primary risk is reflexive demand. If enterprise AI adoption is delayed by 1 to 2 years, or if AI features fail to generate the hundreds of billions in new revenue needed to justify the build-out, depreciation costs on these data centers will severely compress their gross margins. A sustained plateau in AI monetization could force these companies into massive asset write-downs on underutilized infrastructure.
Advanced Semiconductor Foundries (TSMC)
Amount of Exposure: TSMC handles the advanced fabrication and packaging (like CoWoS) for nearly all leading AI chips. The company scaled its 2026 capital investment to approximately $56 billion to keep pace with projected AI silicon demand.
Critical Risk Scenario: Foundries operate on massive fixed costs. If hyperscaler demand cools, leading to a 15% to 20% order cancellation from key designers like Nvidia or AMD, TSMC would be left with idle advanced-node capacity. This overcapacity would compress their profit margins rapidly and jeopardize the return on their heavy 2026 capital outlay.
Server and Infrastructure Assemblers (Super Micro Computer, Dell, HP Enterprise)
Amount of Exposure: These companies build the physical architecture, servers, and liquid cooling systems that house AI GPUs. A significant portion of their forward revenue guidance is reliant on the continuous, uninterrupted rollout of high-density AI server racks.
Critical Risk Scenario: Assemblers operate on notoriously thin margins and are highly vulnerable to inventory risks. If hyperscaler or enterprise demand falls short by just 10% to 15%, they could be left holding expensive, rapidly depreciating hardware components. Because they lack the pricing power of chip designers, even minor delays in deployment can wipe out their net profit margins entirely.
Pure-Play AI Foundation Models (OpenAI, Anthropic)
Amount of Exposure: While private, these companies are the center of the AI narrative. Their exposure is defined by massive cash burn; training state-of-the-art models costs billions of dollars, and daily inference computing costs are astronomical.
Critical Risk Scenario: They are entirely reliant on continuous venture capital and Big Tech funding to survive. If enterprise software revenue misses projections by 30% to 40%, they will struggle to secure the next waves of funding required to train future models. This would lead to severe valuation "down rounds" or potential insolvency if external funding for computing power dries up.
References:
Amiti, M., Redding, S. J., & Weinstein, D. E. (2019). The impact of the 2018 tariffs on prices and welfare. Journal of Economic Perspectives, 33(4), 187–210. https://doi.org/10.1257/jep.33.4.187
Anderson, T. (2026, March 31). Anthropic admits Claude code users hitting usage limits “way faster than expected.” The Register. https://www.theregister.com/2026/03/31/anthropic_claude_code_limits/
Anthropic. (2025, October 23). Expanding our use of Google Cloud TPUs and services. https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services
Anthropic. (n.d.). Anthropic invests $50 bn in American AI infrastructure. https://www.anthropic.com/news/anthropic-invests-50-billion-in-american-ai-infrastructure
The beginner’s guide to Tesla. (2021, May 20). Top Gear. https://www.topgear.com/car-news/electric/beginners-guide-tesla
Bessent, S. (2024). Economic sovereignty and the realignment of the global trading system. [White Paper/Interview series with Wall Street Journal/Bloomberg].
Bloomberg, L. W. &. I. L. /. (2026, March 31). Nasdaq speeds up index entry for SpaceX, large IPOs with new rule. Yahoo Finance. https://sg.finance.yahoo.com/news/nasdaq-speeds-index-entry-spacex-074747727.html
Carnegie Investment Counsel. (n.d.). The risks facing OpenAI and its $1.4T in spending commitments. Carnegie Blog. https://blog.carnegieinvest.com/the-risks-facing-openai-and-its-1.4t-in-spending-commitments
Cavallo, A., Gopinath, G., Neiman, B., & Tang, J. (2019). Tariffs and cost-push inflation (Working Paper No. 26359). National Bureau of Economic Research. https://www.nber.org/papers/w26359
The Decoder. (2026, February 21). OpenAI adds $111 bn to its cash burn forecast as AI costs spiral beyond projections. https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/
Digiday. (n.d.). Media briefing: What The Washington Post’s deal with OpenAI says about the future of AI content licensing. https://digiday.com/media/media-briefing-what-the-washington-posts-deal-with-openai-says-about-the-future-of-ai-content-licensing/
Ding, E. (2025, July 31). tokens are getting more expensive. mandates. https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed
FRED. (2026). Moody’s Seasoned Baa Corporate Bond Yield. https://fred.stlouisfed.org/series/BAA
FRED. (2026a). Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity, Quoted on an Investment Basis (DGS10). https://fred.stlouisfed.org/series/DGS10
Futurism. (n.d.). OpenAI compensation. https://futurism.com/artificial-intelligence/openai-compensation
Google Finance. (2026). Tesla Inc (TSLA) stock price & news - Google Finance. https://www.google.com/finance/quote/TSLA:NASDAQ?hl=en
Hill, B. M. (2026, January 21). Global AI spending will total $2.5 trillion in 2026, says Gartner | Process Excellence Network. Process Excellence Network. https://www.processexcellencenetwork.com/ai/news/global-ai-spending-will-total-25-trillion-in-2026-says-gartner#:~:text=Building%20AI%20foundations%20alone%20will,total%20AI%20spending%2C%20Gartner%20stated.
Internal Revenue Service. (2023). IRS announces delay in Form 1099-K reporting threshold for third-party platform payments for 2023; plans for $5,000 threshold in 2024 (IR-2023-221). https://www.irs.gov/newsroom/irs-announces-delay-in-form-1099-k-reporting-threshold-for-third-party-platform-payments-for-2023
Irwin, D. A. (2011). Peddling protectionism: Smoot-Hawley and the Great Depression. Princeton University Press.
IShares Core U.S. Aggregate Bond ETF | AGG. (n.d.). BlackRock. https://www.ishares.com/us/products/239458/ishares-core-total-us-bond-market-etf#:~:text=52%20WK:%2096.83%20%2D%20101.37,equity%20diversification%20and%20pursue%20income.
Jolly, J. (2020, January 3). Tesla delivers first China-made cars from $5bn Shanghai factory. The Guardian. https://www.theguardian.com/technology/2019/dec/30/tesla-delivers-first-china-made-cars-from-shanghai-factory
Koetsier, J. (2025). Anthropic: Huge Pricing Issues With Glitching Claude Code Limits. Forbes. https://www.forbes.com/sites/johnkoetsier/2026/03/26/anthropic-huge-pricing-issues-with-glitching-claude-code-limits/
Kolodny, L., & Levy, A. (2026, March 13). Elon Musk says xAI must be “rebuilt” as co-founder exodus continues, SpaceX IPO awaits. CNBC. https://www.cnbc.com/2026/03/13/elon-musk-xai-co-founders-spacex-ipo.html
Koparkar, S., & Koparkar, S. (2026, April 3). Leading inference providers achieve lowest token cost with open source models on NVIDIA Blackwell. NVIDIA Blog. https://blogs.nvidia.com/blog/inference-open-source-models-blackwell-reduce-cost-per-token/
KSRED. (n.d.). Claude Code pricing guide: Which plan actually saves you money? https://www.ksred.com/claude-code-pricing-guide-which-plan-actually-saves-you-money/
Lincicome, S. (2024). The limits of presidential power under IEEPA and Section 122. Cato Institute. https://www.cato.org/policy-analysis/limits-presidential-power-trade
Lipshultz, B. (2025, March). OpenAI Expects Revenue Will Triple to $12.7 bn This Year. Bloomberg. https://www.bloomberg.com/news/articles/2025-03-26/openai-expects-revenue-will-triple-to-12-7-billion-this-year
Marshall, A. (2019, February 20). Elon Musk promises a really truly Self-Driving Tesla in 2020. WIRED. https://www.wired.com/story/elon-musk-tesla-full-self-driving-2019-2020-promise/
Mccalip, A. (2025). Economics of Orbital vs Terrestrial Data Centers. https://andrewmccalip.com/space-datacenters
Miran, S. (2024). A User’s Guide to Restructuring the Global Trading System. Hudson Bay Capital. https://www.hudsonbaycapital.com/documents/FG/hudsonbay/research/638199_A_Users_Guide_to_Restructuring_the_Global_Trading_System.pdf
Miran, S. (2024, May 14). The dollar is too strong for America’s own good. Bloomberg Opinion. https://www.bloomberg.com/opinion/articles/2024-05-14/the-dollar-is-too-strong-for-america-s-own-good
MLQ.ai. (n.d.). OpenAI achieves over $20 bn annual revenue run rate amid explosive growth. https://mlq.ai/news/openai-achieves-over-20-billion-annual-revenue-run-rate-amid-explosive-growth/
Morris, R. (2026, April 1). Claude Code users hitting usage limits “way faster than expected.” BBC News. https://www.bbc.com/news/articles/ce8l2q5yq51o
Muppidi, S. (2026, January 22). Anthropic lowers gross margin projection as revenue skyrockets. The Information. https://www.theinformation.com/articles/anthropic-lowers-profit-margin-projection-revenue-skyrockets
NAHC.io Team. (2025, September 23). Anthropic salary overview: How much do employees get paid? NAHC.io. https://www.nahc.io/blog/anthropic-salary-overview-how-much-do-employees-get-paid
New Constructs. (2020, August 20). Tesla: The most dangerous stock for 2020. New Constructs. https://www.newconstructs.com/tesla-the-most-dangerous-stock-for-2020/
News Corp. (2024, May 22). News Corp and OpenAI sign landmark multi-year global partnership. https://investors.newscorp.com/news-releases/news-release-details/news-corp-and-openai-sign-landmark-multi-year-global-partnership
NVIDIA. (2026). Form 10-K filing id: 19184805 Annual Report. https://investor.nvidia.com/financial-info/sec-filings/default.aspx
PYMNTS. (2026, February 18). Anthropic set to pay cloud partners $80 bn through 2029. https://www.pymnts.com/artificial-intelligence-2/2026/anthropic-set-to-pay-cloud-partners-80-billion-through-2029/
Reuters. (2023). Elon Musk’s SpaceX turns profit in first quarter as revenue soars - WSJ. https://www.reuters.com/business/elon-musks-spacex-turns-profit-first-quarter-revenue-soars-wsj-2023-08-17/
Reuters, T. (2026, March 30). New Nasdaq rules to include ‘fast entry’ for new listings on benchmark index. The Mighty 790 KFGO | KFGO. https://kfgo.com/2026/03/30/new-nasdaq-rules-to-include-fast-entry-for-new-listings-on-benchmark-index/
Rezek, R., & Heher, E. (2019). Why fears over the rise in triple-B corporate bonds look overdone. Schroders. https://www.schroders.com/en/malta/professional/insights/why-concerns-over-the-rise-of-triple-b-corporate-bonds-look-overdone/
Sacra. (2026, March 17). OpenAI. https://sacra.com/c/openai/
Sacra. (2026, March 24). Anthropic. https://sacra.com/c/anthropic/
Sam, C., Dottle, R., Ghosh, A., & Kim, K. (2026, March 11). A Guide to the Circular Deals Underpinning the AI Boom. Bloomberg. https://www.bloomberg.com/graphics/2026-ai-circular-deals/
Schenk, C. R. (2010). The decline of sterling: Managing the retreat of an international currency, 1945–1992. Cambridge University Press.
Schiffer, Z., & Matsakis, L. (2025, March 28). Elon Musk’s XAI acquires X, because of course. WIRED. https://www.wired.com/story/xai-x-acquisition-deal/
Silicon Republic. (n.d.). OpenAI secures $4bn in revolving credit. https://www.siliconrepublic.com/business/openai-4bn-credit-bank
Singh, S. (2026, March 25). ChatGPT statistics 2026. DemandSage. https://www.demandsage.com/chatgpt-statistics/
Solving the core Fixed-Income conundrum | Guggenheim Investments. (n.d.). https://www.guggenheiminvestments.com/perspectives/portfolio-strategy/solving-the-core-fixed-income-conundrum-2025
S&P Global. (2026, March). Default, Transition, and Recovery: 2025 Annual Global Corporate Default And Rating Transition Study. https://www.spglobal.com/ratings/en/regulatory/article/default-transition-and-recovery-2025-annual-global-corporate-default-and-rating-transition-study-s101673333
Stan, A. M. (2026, March 28). All 11 xAI co-founders have now reportedly left Elon Musk’s AI company. TNW | Artificial-Intelligence. https://thenextweb.com/news/xai-all-cofounders-departed-musk-spacex-rebuild
Theo - t3․gg. (2025, August 18). I was wrong about AI costs (they keep going up) [Video]. YouTube. https://www.youtube.com/watch?v=mRWLQGMGY80
Thinking Ahead Institute. (2025). Global Pension Assets Study | 2025. https://drive.google.com/file/d/1hyI_50Q6fsCYM-oyRLFYhZVEZsx-6kQO/view?usp=drive_link
Trade Act of 1974, 19 U.S.C. § 2132 (Section 122: Balance of payments authority).
Triffin, R. (1960). Gold and the dollar crisis: The future of convertibility. Yale University Press.
U.S. Department of State. (n.d.). The Smoot-Hawley Tariff Act of 1930. Office of the Historian. https://history.state.gov/milestones/1921-1936/export-import-bank
VanDerhei, J. (2009). The Impact of the Recent Financial Crisis on 401(k) Account Balances. Employee Benefit Research Institute. https://www.ebri.org/content/the-impact-of-the-recent-financial-crisis-on-401(k)-account-balances-4192
Wang, L., & Lee, I. (2026, March 31). Nasdaq speeds up index entry for SpaceX, large IPOs with new rule. Bloomberg. https://www.bloomberg.com/news/articles/2026-03-30/nasdaq-clears-way-for-spacex-big-ipos-to-gain-fast-index-entry
Yeung, K. (2026, April 1). Something is wrong with Claude’s token limits. The AI Economy. https://theaieconomy.substack.com/p/something-wrong-claude-token-limits
York, E. (2026, April 2). Tariff Tracker: 2026 Trump tariffs & trade War by the numbers. Tax Foundation. https://taxfoundation.org/research/all/federal/trump-tariffs-trade-war/
Zitron, E. (n.d.). Where is OpenAI's money going? Where's Your Ed At. https://www.wheresyoured.at/where-is-openais-money-going/
Zuckerman, G., & Ip, G. (2024, October 13). The man who would be Treasury Secretary: Scott Bessent’s plan to remake the economy. The Wall Street Journal. https://www.wsj.com/economy/central-banking/scott-bessent-trump-treasury-secretary-plan-73e44535
Comments
Post a Comment