NVIDIA (NVDA) Deep Dive: Shares Near a Record High, Yet the P/E Is Just 22x—What Is the Market Afraid Of?

🔍 Four-Filter Quick Reference

🎯 Overall Assessment: Cash-equity perspective: “Pass (Build in Tranches)” / Option sellers: “Wait-and-See” — IV has collapsed into the lower end of its 52-week range, making the current risk-reward for collecting premium extremely poor

To state the conclusion upfront: this is ProfitVision LAB's initiation of coverage on NVIDIA. Beginning research only after the AI narrative has run for three years and NVIDIA's market capitalization has exceeded $5.5 trillion has one advantage: we no longer need to prove that AI demand exists. The disadvantage is that all the easy money has already been made. This article therefore does not spend time arguing that “AI is important.” It moves directly to three questions where genuine consensus has yet to emerge: Does the moat reside at the system level or the chip level? Are competitors taking market share or profits? And can the cash behind this impressive income statement actually be collected?

Chapter 1: Industry Map — From Selling Chips to Selling “an Entire Factory”

Understanding NVIDIA's current position requires acknowledging one fact: it is no longer a GPU company. Of its $96.2 billion in FY27Q2 revenue, Data Center contributed $89.0 billion, or approximately 92%. An increasingly large share of that $89.0 billion came not from individual chips, but from complete rack-scale systems—GPU, CPU, NVLink switches, InfiniBand or Spectrum-X networking, rack power, and cooling designs shipped together.

This transition has altered the power structure across the entire industry chain. In the past, hyperscalers designed their own data centers and procured components from different suppliers. NVIDIA now delivers a complete “reference design + full rack” unit, leaving customers to install it in the data center and connect power. Jensen Huang calls this an AI Factory. Beyond the marketing language, the financial implications are direct: order value per customer unit increases by 5x to 10x, while NVIDIA's share of value capture rises rather than falls.

Notably, NVIDIA changed its segment disclosures beginning in FY27Q1. It no longer separately reports Gaming, Professional Visualization, and Automotive, instead consolidating operations into two major segments: Data Center, subdivided into Hyperscale and AI Clouds/Industrial & Enterprise, and Edge Computing. For investors, this change is a double-edged sword. The benefit is that the company has finally acknowledged, “We are a data-center company,” and begun disclosing customer structure. The drawback is that standalone visibility into gaming and automotive has disappeared. If a consumer-market inflection occurs, outside investors will see it later.

The segment within this “other half” most deserving of separate attention is neoclouds—new cloud operators such as CoreWeave and Nebius that specialize in GPU rental. NVIDIA does more than sell them cards; it also invests in them. NVIDIA acquired approximately 9% of CoreWeave for $2 billion and purchased $2 billion of prefunded warrants in Nebius. Yet one frequently overlooked ratio matters: CoreWeave's 2026 capital spending reached $35 billion, while NVIDIA's investment represented only 5.7% of CoreWeave's annual capital spending. NVIDIA's capital is not sustaining these customers; it is providing ignition—using modest equity investments to help create an entirely new customer tier.Demand from this tier is real: CoreWeave has a $104 billion backlog, while Nebius expects more than $9 billion in customer prepayments in 2026.

Regarding market size, rather than citing inconsistent TAM estimates, a more rigorous framework is preferable: NVIDIA's addressable market equals global AI-compute capital spending minus the portion captured by internally developed chips.Using the four major operators' $725 billion of 2026 capital spending as a baseline, approximately 55–65% is directed toward compute and networking hardware, with the remainder allocated to land, power, mechanical and electrical systems, and buildings. After deducting internally developed solutions such as Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia, NVIDIA's accessible pool is approximately $300–350 billion. Compared with the company's own FY27Q3 revenue guidance of $108 billion, equivalent to approximately $432 billion annualized and including non-hyperscaler customers, two conclusions emerge: first, NVIDIA's revenue already exceeds the pool represented by the four major operators, confirming that contributions from “other customers” are both real and substantial; second, the next phase of growth must come from sovereign AI, enterprises, and neoclouds rather than further spending increases by the four major operators.

Chapter 2: Business Model and Moat — CUDA Is the Admission Ticket; NVLink Is the Lock-In

The market's standard answer when discussing NVIDIA's moat is CUDA. That answer is correct, but outdated.

CUDA's moat was strongest from 2020 through 2023. At the time, model-training frameworks, kernels, and optimization toolchains were built almost entirely around CUDA, and switching to another vendor's hardware required engineering teams to rewrite code across the stack. By 2026, however, this barrier had eroded faster than most investors assumed. PyTorch's hardware-abstraction layer has matured, Triton and MLIR have decoupled kernel development from specific vendors, and hyperscalers already write their own internal inference stacks. For these highly capable companies, the incremental cost of switching to TPU or Trainium is no longer high.

The true moat has therefore migrated from the software layer to the system layer. During the FY27Q2 earnings call, management quantified this transition with a powerful metric: revenue opportunity per GW of data-center capacity.

For the same 1GW AI data center, NVIDIA's revenue opportunity has more than doubled within two generations. This is the most direct financial evidence of its moat: even if internally developed ASICs continue to divert orders, NVIDIA's revenue growth can significantly exceed customer capital-spending growth as long as its value capture per GW continues to rise.Conversely, the day this figure stops increasing will mark the first evidence that the moat has begun to leak.

At the system level, NVIDIA's competitive advantages can be divided into four layers, ranked from strongest to weakest:

Reading this table in reverse produces an uncomfortable conclusion for bulls: NVIDIA's strongest moat is also its most capital-intensive and upstream-dependent layer.Its system-integration advantage rests on priority access to CoWoS advanced packaging and HBM4 capacity. If memory supply loosens or TSMC makes packaging capacity more broadly available, this advantage will be diluted. Current scarcity benefits NVIDIA, but scarcity is not a permanent business model; it is a cycle.

Specific Scenarios in Which the Moat Could Be Breached

An honest assessment requires a falsifiable scenario rather than a vague statement that “competition will intensify.” The most probable path is as follows:

Step One (already occurring): Inference workloads exceed training workloads. Inference requires far less interconnect bandwidth than training, materially reducing NVLink's value in inference environments. Yet inference represents the largest long-term revenue pool.

Step Two (currently occurring): Hyperscalers dedicate internally developed ASICs to inference. Google's seventh-generation TPU, Ironwood, was formally launched in April 2026. Broadcom confirmed that Anthropic will deploy 1GW of TPU capacity in 2026 and expand to 3GW in 2027. Microsoft Maia 200 and AWS Trainium3 are also designed around inference economics.

Step Three (not yet occurring and the critical monitoring point): One hyperscaler publicly announces that internally developed chips will serve as the primary platform for incremental inference capacity, while NVIDIA will be used only for training. Once this occurs, NVIDIA's growth curve will be downgraded from tracking total capital spending to tracking training spending, which grows far more slowly than inference spending.

We are currently in the middle of Step Two. However, the market broadly underestimates one development: NVIDIA did not choose to confront ASICs using GPUs alone. It acquired the strongest non-GPU inference architecture outright.

Countermeasure One: $20 Billion to Acquire Groq's Inference Technology and Team

In December 2025, NVIDIA completed an asset transaction involving inference-chip company Groq for approximately $20 billion, the largest transaction in NVIDIA's history. The structure deserves close attention. NVIDIA did not acquire the entire company; the transaction consisted of a non-exclusive technology license plus an asset acquisition. Groq founder Jonathan Ross, President Sunny Madra, and the core team joined NVIDIA, while Groq continued to operate independently under newly appointed CEO Simon Edwards.

The product entered full-scale production in August 2026. The Groq 3 LPX rack achieves decoding throughput of 3,400 tokens per second. When paired with Vera Rubin NVL72 to run trillion-parameter models, NVIDIA claims token output per watt can reach up to 35x that of Blackwell NVL72.

The strategic significance far exceeds the immediate financial contribution: when hyperscalers use internally developed ASICs to attack NVIDIA's weakness in inference, NVIDIA's answer is not merely, “My GPUs can also perform inference.” It is, “I also offer a dedicated inference machine, integrated with my GPU racks through the same NVLink and software stack.”This reframes the competition from “GPU vs. ASIC” to “complete heterogeneous-computing system vs. point accelerator.” Systems are precisely where NVIDIA's moat is deepest.

Countermeasure Two: Vera CPU Opens an Entirely New $200 Billion Market

The second front receives even less attention. Vera is not merely a supporting CPU within Vera Rubin racks. NVIDIA positions it as the world's first CPU designed for agentic AI, and it is also sold separately. Jensen Huang estimates the market opportunity at $200 billion, explicitly including China. Standalone Vera CPU revenue is expected to reach $20 billion in FY2027, with initial customers including Anthropic, OpenAI, and Oracle Cloud.

The logic is straightforward: models “think” on GPUs, but when AI agents execute tasks—calling tools, reading and writing files, and connecting systems—the majority of workloads run on CPUs. As agentic AI scales, the primary requirement is not simply more GPUs, but more CPUs optimized for these workloads.

The competitive implications are two-sided: for NVIDIA, this represents expansion of the revenue pool; for Intel and AMD, it represents a flanking attack on their core business. In the fourth quarter of 2025, before Vera's launch, server CPU shipment share was 60% for Intel, 24.3% for AMD, and 6.2% for NVIDIA. Counterpoint expects Arm architecture to capture 90% of the server CPU market by 2029.

Returning to gross margin, both strategic fronts require NVIDIA to improve inference efficiency enough that the cost advantage of internally developed ASICs cannot offset the ecosystem gap. This is technically feasible, but the cost is higher R&D investment and product-mix dilution. Gross margin is therefore the subject of the next chapter.

Chapter 3: Competitive Landscape — Many Can Take Share; Few Can Take Profits

For the first time in 2026, the AI accelerator market has reached a point where alternatives to NVIDIA can genuinely ship at scale. Yet a closer examination shows that the nature of each competitive threat differs substantially.

Who Is the Real Threat?

The answer is not AMD. It is Broadcom.

The reasoning is as follows: AMD and NVIDIA compete for the same budget, whereas Broadcom changes how the budget is allocated.When Microsoft evaluates whether to buy MI455X or Rubin, the result is a pricing contest. NVIDIA may earn less, but the market structure remains unchanged. When Google decides that internally developed TPUs will support incremental inference capacity, that spending never enters the external accelerator market. It becomes an internal Google cost and a Broadcom design-services fee.

Analyst estimates place NVIDIA's 2026 share of AI training accelerators above 80%, with AMD at approximately 5–7%. This is positive for bulls, but the metric measures training and the merchant market, excluding the two dimensions where the threat is greatest: inference and internally developed chips. The correct metric to monitor is not market share, but the difference between NVIDIA Data Center revenue growth and capital-spending growth among the four major hyperscalers. In 2026, this spread was positive: NVIDIA grew 117% while hyperscaler capital spending grew 77%. This indicates that NVIDIA was not being diluted; it was gaining share from others. The first structural inflection will occur when this spread turns negative.

Market share can mislead because it measures which battles were won. The growth-rate spread cannot, because it measures whether the entire war is being won.

AMD also presents an often-overlooked threat: it does not need to win; it only needs to exist.Even if MI455X never reaches 15% market share, specifications sufficiently close to Rubin give every procurement executive negotiating leverage. This does not appear in financial statements as immediate share loss, but as gradual gross-margin erosion. During the FY27Q2 earnings call, NVIDIA explicitly provided a complete gross-margin trajectory:

The company cited “extreme pricing conditions” in memory, with increases exceeding prior expectations and further increases expected next year. Its countermeasure is to raise prices for GB and Vera Rubin server systems by more than 15% beginning in early 2027. Jensen Huang confirmed on the earnings call that the increases would take effect early next year.

The critical issue is not 71%, but the causal sequence: a recovery to 72–73% is not a forecast; it is a conditional statement. Whether the price increase can be fully passed through to Microsoft, Google, and Oracle will determine whether FY2028 gross margin is 72% or 68%.

In other words, memory costs are a valid explanation, but not the complete explanation. Gross margin is being compressed simultaneously by upstream price increases and downstream bargaining pressure, and NVIDIA has chosen to pass those costs through. Whether this pass-through succeeds is where the threat posed by AMD's mere existence will be quantified.

Chapter 4: Financial Resilience — The Income Statement Is Flawless; the Balance Sheet Is Beginning to Speak

First, the good news—and it is exceptionally good.

Quarterly net income of $59.7 billion and a 62% net margin are nearly unprecedented in the history of large public companies. Operating cash flow was $24.1 billion, free cash flow was $21.3 billion, and nearly $26.0 billion was returned to shareholders in a single quarter through repurchases and cash dividends. Cash and marketable securities totaled $56.6 billion. By any conventional standard, this income statement earns full marks.

The value of research, however, does not lie in repeating the perfect score. It lies in determining how that score was calculated. Three signals warrant caution.

Signal One: Days Sales Outstanding Jumped from 45 Days to 60 Days

FY27Q2 accounts receivable reached $63.1 billion, while days sales outstanding, or DSO, increased from 45 days in the prior quarter to 60 days. The company explained that it had provided extended payment terms under large, multi-quarter contracts with certain investment-grade customers.

More important than the number of days is the simultaneously disclosed concentration. The 10-Q disclosed that the top five direct customers represented 22%, 14%, 13%, 11%, and 10% of accounts receivable, respectively, for a combined 70%. One year earlier, the figure was only 56%, representing an increase of 14 percentage points within 12 months.

The implication is clear: the customers receiving relaxed payment terms are the same customers making the largest revenue contributions.This is not routine working-capital expansion distributed across hundreds of customers. The payment cadence of a small number of very large customers directly determines NVIDIA's cash flow.

Signal Two: GAAP EPS Exceeded Non-GAAP EPS

FY27Q2 GAAP diluted EPS was $2.46, compared with non-GAAP EPS of $2.22. This ordering is unusual in NVIDIA's history, because non-GAAP EPS typically exceeds GAAP EPS after excluding stock-based compensation.

The most plausible explanation is that GAAP net income included non-operating investment valuation gains. In other words, changes in the fair value of NVIDIA's equity investments in AI companies such as OpenAI, Anthropic, and xAI were recognized in GAAP earnings and excluded from non-GAAP results. If this explanation is correct, then a portion of the $59.7 billion in GAAP net income was not earned from selling chips, but from increases in the valuations of customers in which NVIDIA had invested.

★ This item has been verified: the inference is correct.The 10-Q reconciliation disclosed net gains on equity securities of $7.8 billion for the quarter and unrealized investment gains of $12.5 billion at period-end. This confirms that GAAP net income exceeded non-GAAP net income primarily because these investment valuation gains were recognized under GAAP and excluded from non-GAAP results. Regardless of the precise allocation, investors should use non-GAAP EPS of $2.22 rather than GAAP EPS of $2.46 when assessing core operating profitability.

Signal Three: Supply Commitments of $279 Billion and Inventory Rising from $25.8 Billion to $31.6 Billion

The company explained the inventory increase as preparation for the Vera Rubin introduction in the third quarter, which is entirely reasonable during a product-generation transition. However, $279 billion in supply commitments, most of which relate to memory required for Rubin, is a different matter. These represent non-cancelable purchase obligations or commitments with high termination costs, equivalent to approximately seven quarters of future cost of revenue.

The significance is straightforward: NVIDIA has placed its bet.If demand develops as management expects, including approximately 70% FY2028 revenue growth under supply-constrained conditions, these commitments will convert into revenue. If demand abruptly stalls in 2027, they will become impairments. NVIDIA demonstrated this mechanism in FY26Q1, when H20 export restrictions resulted in a $4.5 billion impairment of inventory and purchase obligations. A recurrence against a $279 billion commitment base would be far more damaging.

Capital Structure and Cash Quality Summary

Connecting these items produces the chapter's most important conclusion: NVIDIA's income statement is accelerating while its cash-flow statement is decelerating.Only $21.3 billion of $59.7 billion in net income converted into free cash flow, a 36% conversion rate that is far below the company's historical level. The difference is locked in accounts receivable and inventory. This is not necessarily negative, because working capital naturally expands at rapidly growing companies. It does mean, however, that the quality of today's earnings depends on whether customers can pay tomorrow.The identity of those customers and the source of their funding are the central variables in the next chapter's valuation scenarios.

Chapter 5: Valuation and Scenario Analysis — The Market Has Already Priced In an “Abrupt Halt”

First, establish the baseline. Using the intraday share price of $227.98 on August 27, 2026 and a market capitalization of approximately $5.5 trillion:

Viewed together, the conclusion is clear: the market is not valuing NVIDIA at a bubble multiple.A company with quarterly revenue growth of 106% year over year and a 62% net margin trades at only 21–23x earnings, below the average P/E for the S&P 500 information technology sector. The market believes the current quarter's figures are real, but refuses to believe they can persist for more than two years.

In other words, the current share price already embeds the assumption that AI capital spending will abruptly halt at some point. This changes the nature of the entire investment thesis. Buying NVIDIA is no longer a bet that the AI narrative will continue; that was already reflected in the gains from 2023 through 2025. It is a bet that the abrupt halt will occur later and unfold more gradually than the market expects.

Three-Scenario Analysis

This article does not forecast a target price. The three scenarios below serve one purpose: to clarify the assumptions under which the current price is expensive or inexpensive.

What Is the Market Currently Pricing?

Comparing the three scenario multiples with the current price allows the market's embedded assumptions to be inferred. The current price equates to approximately 16.6x under the base scenario. This multiple is low for a company growing 40%, with a PEG ratio of approximately 0.4, but represents reasonable risk compensation for a company that could enter the bear scenario within two years.

This is NVIDIA's central valuation paradox: it is simultaneously an inexpensive growth stock and an expensive cyclical stock. The classification depends on whether AI capital spending represents a new long-term infrastructure cycle or a one-time arms race.No financial statement can answer that question; only time can.

Our answer is: both.The long-term trend in AI-compute demand is real, but the current spending trajectory is unsustainable. Capital spending by the four major operators rose 77% in 2026, with consensus expecting another approximately 29% increase in 2027. This rate must eventually normalize toward revenue growth. The only question is whether this occurs in 2027 or 2029. That 24-month difference represents the entire distinction between the bull and bear scenarios.

Chapter 6: Conclusion and Tactical Recommendations

Core View

At the current price, NVIDIA offers attractive odds but requires strict position sizing. The risk lies not in valuation, but in whether investors are purchasing peak-cycle earnings.

Options Tactics: This Is Not a Seller's Market

The post-earnings decline in IV has been severe. NVDA's current annualized implied volatility is approximately 33.2%, while its 30-day historical volatility is 35.9%. IV is below HV and stands near the extremely low 4th percentile of its 52-week range.

Accordingly, no strike-price parameters are provided. Investors holding the shares who are considering covered calls to reduce cost basis must choose strikes and expirations according to their individual cost basis, account size, and risk tolerance. However, at the 4th IV percentile, premiums from any seller structure are extremely thin, and the opportunity cost may exceed the income. The honest conclusion is that the best options action for this security at present is no action.

Triggers: Conditions for an Upgrade or Downgrade

The table below defines the monitoring discipline for this research. These upgrade and downgrade rules were established in advance to ensure future updates remain auditable and unaffected by emotion. They describe how this research view will change and do not constitute trading instructions for any reader. Readers should make independent judgments based on their financial circumstances and risk tolerance.

Position Management

Under ProfitVision LAB's 5% risk-unit framework, NVIDIA's fundamental volatility, reflected in 106% earnings growth, and price volatility, with HV30 at 35.9%, are both substantially higher than those of typical large-cap stocks. The company is also highly correlated with the entire AI supply chain. If a portfolio already holds positions related to TSMC, Broadcom, memory, or cloud services, actual risk exposure is materially greater than the stated portfolio weight. Accordingly, this research adopts the following position in its tracking record: size exposure at the lower bound, rather than the upper bound, of one risk unit, and manage aggregate AI supply-chain exposure as a single position. Whether to apply this framework and how to allocate capital should be evaluated according to each investor's financial circumstances.

📋 Tracking Record

Next expected update:After FY2027 third-quarter earnings, expected in mid-to-late November 2026. Priority review items, in order, are: ① whether DSO declines and the top five customers' share of accounts receivable stops rising; ② whether non-GAAP gross margin meets 74% guidance and whether Q4 trough guidance is reduced again; ③ whether Vera Rubin reaches management's guidance of approximately 20% of revenue; ④ actual Groq 3 LPX shipments and customer adoption, and whether standalone Vera CPU revenue remains on track toward approximately $20 billion in FY2027.

Conditions triggering an early update:① any hyperscaler reduces 2027 capital-spending guidance; ② NVIDIA announces any inventory or purchase-commitment impairment; ③ terms of the SB Energy/OpenAI guarantee change, or provisions disclosed in the 10-Q exhibits differ from the current understanding; ④ the $500 billion financing platform makes material progress or breaks down, including identification of the first project, disclosure of a committed amount by any partner, publication of collateral terms, or a delay to the platform; ⑤ customers publicly resist or delay the price increase of more than 15% planned for early 2027; ⑥ U.S. AI-chip export policy toward China changes materially again; ⑦ the share price falls below the 200-day moving average, currently approximately $195.73.

Frequently Asked Questions

Q: What is NVIDIA's primary business?

NVIDIA has transformed from a graphics-chip company into an AI data-center systems provider. Of its $96.2 billion in revenue for the second quarter of fiscal 2027, ended July 2026, Data Center contributed $89.0 billion, or approximately 92%. NVIDIA delivers more than GPUs; it supplies full-rack systems incorporating GPUs, CPUs, NVLink high-speed interconnects, InfiniBand, and Spectrum-X networking. Its major customers fall into two categories, each representing approximately half of Data Center revenue: hyperscale cloud operators such as Microsoft, Amazon, Google, and Meta; and AI cloud providers, sovereign AI projects, industrial customers, and enterprises. Beginning in the first quarter of fiscal 2027, the company simplified its reporting segments into Data Center and Edge Computing.

Q: What is the difference between Blackwell and Vera Rubin?

They are two consecutive generations of NVIDIA's AI-computing platform. Blackwell, including the updated Blackwell 300, is the current primary shipping product and drove the revenue surge in the first half of 2026. Vera Rubin is the successor platform and entered full-scale production in the second half of 2026. Management described it as the fastest product ramp in company history and expects it to contribute approximately 20% of third-quarter revenue. Technically, Rubin uses HBM4 memory and the new Vera CPU to deliver a substantial improvement in inference efficiency, which is central to NVIDIA's defense against internally developed ASICs. NVIDIA also holds two less-discussed assets: the Groq 3 LPX dedicated inference rack introduced after the approximately $20 billion acquisition of Groq's inference technology and team in December 2025, which entered full-scale production in August 2026; and the standalone Vera CPU, expected to generate approximately $20 billion in FY2027 revenue. Management's objective is for Blackwell and Rubin to generate approximately $1 trillion in combined revenue from 2025 through the end of 2027.

Q: Is NVIDIA currently profitable, and what is its financial condition?

Profitability is exceptionally strong. In the second quarter of fiscal 2027, GAAP net income reached $59.7 billion, up 126% year over year, with a net margin of approximately 62% and a GAAP gross margin of 75.0%. Cash and marketable securities totaled $56.6 billion, quarterly free cash flow was $21.3 billion, and nearly $26.0 billion was returned to shareholders. Two points require attention. First, quarterly GAAP EPS of $2.46 exceeded non-GAAP EPS of $2.22. This was verified as primarily resulting from $7.8 billion in net gains on equity securities, with period-end unrealized gains of $12.5 billion; core operating performance should therefore be assessed using non-GAAP results. Second, free cash flow represented only 36% of GAAP net income, with the difference tied up in accounts receivable of $63.1 billion, where DSO increased from 45 days to 60 days, and inventory of $31.6 billion.

Q: What is the greatest investment risk facing NVIDIA?

The core risk is the combination of circular financing and collection quality. NVIDIA has proposed an investment framework of up to $100 billion for OpenAI and, in August 2026, provided a credit guarantee capped at $105 billion for an Ohio data-center campus leased by OpenAI, producing maximum gross exposure of approximately $108.5 billion when combined with existing partner support. Across the AI ecosystem, cross-holdings have reached approximately $46 billion and multi-year purchase commitments approximately $879 billion. This means part of NVIDIA's revenue comes from customers it finances or guarantees. The trigger mechanism is as follows: if end-market paid revenue from AI applications grows more slowly than expected, customers first delay payments—DSO has already risen from 45 days to 60 days, while the top five customers' share of accounts receivable increased from 56% to 70% within one year. Customers then cancel or defer orders, while NVIDIA cannot simultaneously cancel its $279 billion in upstream supply commitments, eventually converting those commitments into inventory impairments. The company demonstrated this mechanism in the first quarter of fiscal 2026, when H20 export restrictions led to a $4.5 billion impairment. The second risk is that internally developed ASICs from hyperscalers, including Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia, move inference demand outside the merchant market.