How this report was made — CUDA moat migrated to silicon and will persist through 2028

This is the complete methodology record for the report above: what we asked, what we read, what we attacked, what survived, and — just as importantly — what we never got to test. It is published in full so you can judge how much weight the conclusion bears.

Report state: published · Verification: verified Published: 2026-08-11T19:15:48.993+00:00 · Record generated: 2026-08-11T19:15:49.559Z

1. The question

Question: NVDA: Is CUDA still a moat, for how long and is it in the price? Subject: NVDA Angle: Three reasons, none of which is "another NVIDIA note". First, the debate is being fought with 2023 arguments on both sides. Bulls cite five million CUDA developers; bears cite PyTorch portability. Both describe a world that no longer exists: CUDA-the-API is already largely bypassed, and yet the moat did not die — it moved. Where it moved to, and how durable each layer is, is the actual question. Second, the moat question and the stock question have decoupled, and almost nobody models them separately. History's clearest lesson — Microsoft, Cisco, Intel, Oracle — is that platform moats survive while their stocks deliver dead-money decades. We assess them as two questions, not one. Third, this report was built adversarially. Every seat's claims were attacked by independent verifiers with live source-checking before publication. Several claims died. Some of our own probabilities were found incoherent and were reconciled. The annex shows the corrections — a research product that displays its own error bars is, we believe, worth more than one that hides them.

2. Thesis, antithesis and how it can be proved wrong

Thesis: NVIDIA's moat has already shifted from CUDA-the-API to hardware-software co-design (NVLink, Tensor Cores, NeMo) that AI agents cannot replicate, creating $15–25 million switching costs per 10,000-GPU cluster that will sustain 70%+ gross margins and justify a $320–350 fair value through the Vera Rubin cycle ending fiscal 2028. Antithesis: Hyperscaler dual-sourcing to AMD and custom ASICs will capture 30–40% of inference capacity by end of 2027, capping NVIDIA data center growth at 15–20% annually and keeping the stock range-bound as customer concentration (45–55% in five accounts) forces pricing concessions that compress multiples regardless of moat persistence. Falsifiability event: Q2 FY2027 results, 2026-08-26 [e25] — 2026-08-26 What it would show: Data center gross margin below 70% would signal pricing pressure from hyperscaler negotiating power or AMD competition; revenue growth below 60% year-over-year would indicate material share loss to custom ASICs; customer concentration disclosure showing top-5 dropping below 40% would validate diversification thesis; any guidance cut or margin compression invalidates the vertical integration defense and confirms the market's moat-decay pricing. Why this event: Q2 results arrive in 17 days and will show whether gross margin holds (validating hardware lock-in) or compresses (validating hyperscaler pricing power), making it the earliest falsifiable test of competing hypotheses; the angle notes emphasize adversarial testing and displaying error bars, so we anchor on the nearest hard data rather than waiting for Vera Rubin launch in H2 2026 when multiple narratives will already be entrenched.

3. What we took "now" to mean

Fiscal year: FY2027 · Quarter in progress: Q2 FY2027 Last reported quarter: Q1 FY2027 (ended 2026-04-26) Next expected filing: Q2 FY2027 2026-08-26 [e25] — estimated from filing cadence Established via: the SEC EDGAR submissions API · Confidence: high

Every date-discipline ruling below was decided against that calendar: a claim about "the latest quarter" is judged against the quarter that had actually reported when the research ran, not the calendar quarter.

Filings consulted to fix the calendar

4. Everything we read

39 source(s): 8 T1, 4 T2, 3 T3, 24 T4. T1 is a primary filing or the company itself; T4 is unattributed commentary.

5. The evidence we kept

Superseded during reconciliation (9) — a better-tiered or more recent figure replaced these, so the drafter never saw them.

Figures that primary sources disagree on (1) — the drafter was instructed to cite the filing or omit the figure.

Discarded before drafting: 17 item(s) discarded for failing attribution or date validation. They are not itemised here: the filter is disclosed, individual third-party sources are not named.

5b. Charts

2 of 2 nominated chart(s) shipped.

6. The attack set

Before publication every load-bearing claim in this report was attacked by an adversarial pass. Drafting used Anthropic Claude; verification used Google Gemini. No family appears on both sides, so the check is not the writer marking its own work. Exact model versions are configuration and are not published.

Claims were clustered by substance first: 7 cluster(s) drawn from 11 phrasings. Each cluster drew one ruling; identical phrasings inherit it.

CAVEAT — CUDA moat migrated to silicon and will persist through 2028

Severity: load_bearing · Source tier: T1 · Cluster: c1 Reasoning: The claim overstates the durability of NVIDIA's competitive advantage. While the company is expected to remain dominant, multiple sources forecast a material erosion of its market share through 2028. Celadon Research projects NVIDIA's share will fall to 65-70% by 2028 due to competition from custom silicon and AMD. This contradicts the idea that the moat will 'persist', which implies a steady state of dominance. Because the claim overstates its case based on the available forecasts, the verdict is caveat. Components ruled

Decisive component: The claim that the moat 'will persist through 2028' was decisive. While NVIDIA is projected to remain the market leader, forecasts of its market share declining from over 80% to 65-70% by 2028 represent a significant erosion, not a persistence, of its current competitive advantage. This makes the claim an overstatement. Corroboration: none validated Counter-evidence

NVIDIA's market share in AI accelerators will erode from current levels above 80% toward 65–70% by 2028 as custom silicon from hyperscalers and AMD's MI-series capture incremental inference workloads despite NVIDIA's ecosystem advantages.

While percentage share will decline to 75% by 2026 as AMD and custom silicon scale, NVIDIA's absolute revenue continues to grow because the total addressable market is expanding faster than any single competitor can capture.

llama.cpp by Georgi Gerganov is the most-deployed open-source LLM inference project in the world. The 2026 ecosystem covers GGUF as the de facto cross-platform model format, plus hardware backends for NVIDIA (CUDA), AMD (ROCm + Vulkan), Apple (Metal), Intel (SYCL + Vulkan), Qualcomm (OpenCL), CPU (AVX-512 + ARM Neon), plus downstream projects including

CAVEAT — NVIDIA's moat has already shifted from CUDA-the-API to hardware-software co-design (NVLink, Tensor Cores, NeMo) that AI agents cannot replicate, creating $15–25 million switching costs per 10,000-GPU cluster that will sustain 70%+ gross margins and justify a $320–350 fair value through the Vera Rubin cycle ending fiscal 2028.

Severity: load_bearing · Source tier: T1 · Cluster: c1 Reasoning: The claim overstates the durability of NVIDIA's competitive advantage. While the company is expected to remain dominant, multiple sources forecast a material erosion of its market share through 2028. Celadon Research projects NVIDIA's share will fall to 65-70% by 2028 due to competition from custom silicon and AMD. This contradicts the idea that the moat will 'persist', which implies a steady state of dominance. Because the claim overstates its case based on the available forecasts, the verdict is caveat. Components ruled

Decisive component: The claim that the moat 'will persist through 2028' was decisive. While NVIDIA is projected to remain the market leader, forecasts of its market share declining from over 80% to 65-70% by 2028 represent a significant erosion, not a persistence, of its current competitive advantage. This makes the claim an overstatement. Corroboration: none validated Counter-evidence

NVIDIA's market share in AI accelerators will erode from current levels above 80% toward 65–70% by 2028 as custom silicon from hyperscalers and AMD's MI-series capture incremental inference workloads despite NVIDIA's ecosystem advantages.

While percentage share will decline to 75% by 2026 as AMD and custom silicon scale, NVIDIA's absolute revenue continues to grow because the total addressable market is expanding faster than any single competitor can capture.

llama.cpp by Georgi Gerganov is the most-deployed open-source LLM inference project in the world. The 2026 ecosystem covers GGUF as the de facto cross-platform model format, plus hardware backends for NVIDIA (CUDA), AMD (ROCm + Vulkan), Apple (Metal), Intel (SYCL + Vulkan), Qualcomm (OpenCL), CPU (AVX-512 + ARM Neon), plus downstream projects including

Qualifications: inherited from an identical claim

REFUTED — NVIDIA's moat has already migrated from CUDA-the-API to hardware-software co-design: the $6.9bn Mellanox acquisition in March 2019 secured the InfiniBand stack, and the integration of NVLink, Tensor Cores, and NeMo creates switching costs that AI agents cannot replicate by writing PyTorch wrappers.

Severity: supporting · Source tier: T1 · Cluster: c2 Adversary stated CAVEAT, overruled to REFUTED by its own weakest component Reasoning: The claim's assertion about high switching costs from hardware-software co-design, specifically NVLink, is strongly corroborated by sources [4] and [5], which show performance degradation of 23% to 130x when using alternatives. However, the claim also presents the 2019 Mellanox acquisition and its InfiniBand stack as a core pillar of this moat. Source [2], covering 2023-2026, directly refutes the continued dominance of InfiniBand, stating Ethernet has overtaken it in market share. Because one of the key technological pillars cited for the moat is shown to be losing its competitive leadership, the overall claim is overstated. The weakest link is the InfiniBand assertion, which is refuted, thus the overall verdict is a CAVEAT. Components ruled

Decisive component: The weakest component is the inferred durability of the InfiniBand moat. While the acquisition is a fact and the NVLink integration provides a demonstrable performance advantage, the claim's reliance on InfiniBand as a key part of the *current* moat is undermined by evidence of its declining market share relative to Ethernet. This refutes a key premise of the claim's construction, forcing an overall verdict of CAVEAT. Corroboration

You can see that NVLink completes the training ~23% faster. In the second benchmark we useNCCL_P2P_DISABLE=1 to tell the GPUs not to use NVLink, which will use PCIe instead.

Synchronizing 1 GB of gradients takes about 0.6 ms over NVLink 5, 16 ms over a PCIe link (64 GB/s), and 80 ms over a 100-gigabit Ethernet link — same model, same math, a 130× spread

Counter-evidence

Ethernet is now the leader in “scale-out” AI networking. In 2023, InfiniBand held an ~80% share of the data center switch market. A little over two years later, Ethernet has overtaken it in data center switch and server port counts. Indeed, the demand for Ethernet-based interconnect technologies continues to strengthen, reflecting the market’s broader shift toward scalable, open, and cost-efficient da

CONFIRMED — Two unnamed customers accounted for approximately 13% of NVIDIA's total revenue in Q1 FY2027, period ended April 26, 2026.

Severity: load_bearing · Source tier: T1 · Cluster: c3 Reasoning: The claim asserts that two customers accounted for approximately 13% of revenue in Q1 FY2027. The provided counter-evidence includes NVIDIA's Form 10-Q for the period ended April 26, 2026 (Q1 FY2027). This primary source document explicitly states, 'In the first quarter of fiscal year 2027, two unnamed customers each accounted for 10% or more of our total revenue, and in aggregate, they represented approximately 13% of our total revenue.' As the claim is a direct and accurate statement from the primary financial filing for the specified period, it is confirmed. Components ruled

Decisive component: The claim consists of a single, verifiable fact which was positively corroborated by the primary source document. Corroboration

In the first quarter of fiscal year 2027, two unnamed customers each accounted for 10% or more of our total revenue, and in aggregate, they represented approximately 13% of our total revenue.

Counter-evidence: none

REFUTED — NVIDIA reported gross margin of 78.4% for Q1 FY2027, period ended April 26, 2026, consistent with the 78% data center gross margin for FY2026.

Severity: load_bearing · Source tier: T1 · Cluster: c4 Reasoning: The claim asserts a gross margin of 78.4% for Q1 FY2027. However, NVIDIA's official financial results for the first quarter ended April 26, 2026 (Q1 FY2027) explicitly state that the non-GAAP gross margin was 75.0%. As the counter-evidence is from the company's own press release and covers the exact same period, it directly refutes the factual basis of the claim, making the overall verdict 'refuted'. Components ruled

Decisive component: The decisive component is the claim's primary observed fact: the 78.4% gross margin figure for Q1 FY2027. This is the weakest component because it is a specific, verifiable number that is directly contradicted by primary source evidence for the same period. Corroboration: none validated Counter-evidence

For the quarter, GAAP and non-GAAP gross margins were 74.9% and 75.0%, r

For the quarter, GAAP and non-GAAP gross margins were 74.9% and 75.0%, respectively.

For the quarter, GAAP and non-GAAP gross margins were 74.9% and 75.0%, respectively.

CAVEAT — Cerebras CEO Andrew Feldman claims CUDA has lost approximately 70% of frontier AI training share.

Severity: supporting · Source tier: T4 · Cluster: c5 Reasoning: The claim, attributed to a competitor's CEO, concerns a specific and recent loss of market share in 'frontier AI training'. The counter-evidence provided is for calendar years 2022 and 2023, making it too stale to refute a claim about the market in mid-2026, especially in a rapidly evolving sector like AI. Additionally, the evidence speaks to Nvidia's dominance in the overall training or data center accelerator market, which is not the same as the 'frontier' segment specified in the claim. Therefore, the evidence neither corroborates nor contradicts the claim. Components ruled

Decisive component: The claim's core assertion about a 70% loss of share in 'frontier AI training' could not be tested. All available counter-evidence is from 2022 and 2023, making it too dated to address a claim about the market in mid-2026. Furthermore, the evidence discusses the broader training market, not the specific 'frontier' segment. Corroboration: none validated Counter-evidence

Nvidia remains the market leader in the training market segment with no competition in sight. Intel leads the inference market segment. Nvidia (36%) and Intel (40%) are very close in terms of overall DC AI Chip market share, and Nvidia is expected to become the overall market leader in 2023.

Nvidia is clearly the leader in the market for training chips, but that only makes up about 10%-20% of the demand for AI chips. There is a far larger market for Inference chips, which take trai

Nvidia proved to be king of data center servers and storage in 2023 as artificial intelligence (AI) ballooned in boardroom talks and application deployments, leading to a 224% growth in revenue from GPUs, custom accelerators and FPGA, accordi

Qualifications: sourced on T4 commentary only

REFUTED — NVIDIA's data center revenue grew 92% year-over-year in Q1 FY2027, to $26.3bn from $13.7bn in Q1 FY2026.

Severity: load_bearing · Source tier: T1 · Cluster: c6 Reasoning: While the claim correctly identifies the 92% year-over-year growth rate for NVIDIA's Data Center segment in Q1 FY2027, it provides incorrect absolute revenue figures for both the current and prior-year periods. The company's press releases for Q1 FY2027 and Q1 FY2026 report Data Center revenue of $75.2 billion and $39.1 billion, respectively, directly refuting the claim's figures of $26.3 billion and $13.7 billion. Because the foundational revenue numbers are incorrect, the claim is refuted. Components ruled

Decisive component: The claim's stated revenue figures for Q1 FY2027 ($26.3bn) and Q1 FY2026 ($13.7bn) were the decisive components, as both were directly contradicted by the company's own financial reports for those periods. Corroboration

Record Data Center revenue of $75.2 billion, up 92% from a year ago

Record Data Center revenue of $75.2 billion, up 92% from a year ago

Counter-evidence

Record Data Center revenue of $75.2 billion, up 92% from a year ago

Data Center revenue of $39.1 billion, up 10% from Q4 and up 73% from a year ago

CAVEAT — Over 5.9 million developers worldwide use CUDA and NVIDIA's other software tools as of FY2026, fiscal year ended January 25, 2026.

Severity: supporting · Source tier: T1 · Cluster: c7 Reasoning: The claim makes a specific factual assertion about the number of developers for fiscal year 2026. The counter-evidence includes a source stating there were 'approximately 5.9 million' developers in FY2025. While this number from the prior year does not contradict the claim for FY2026, it also does not corroborate it. No evidence was provided for the specific period of the claim (FY2026). Therefore, the claim is unsupported by the available material. Components ruled

Decisive component: The claim's core assertion about the developer count in FY2026 is the decisive component. It is the weakest link because it is not supported by any of the provided evidence, which only contains data for the prior fiscal year (FY2025). Corroboration: none validated Counter-evidence

Developers using CUDA & NVIDIA software tools (FY2025): approximately 5.9 million. (NVIDIA 2025 Annual Report (PDF))

7. How much weight the challenge can bear

These are the limits of the check above. They are published for the same reason the annex is: a challenge you cannot see the edges of is not a challenge. Never tested: none — every claim in the set drew a ruling.

Rulings cut short: none.

Ruled by the fallback model: none.

Counter-search composed mechanically: none — every claim drew authored adversarial queries.

Resting on T4 commentary only (1).

Sourcing flags raised while outlining (3).

Excluded as unverifiable (3). These assert what a document that does not yet exist will show, so they could not be verified in either direction and were kept out of the attack set.

7b. The house number

The model's central output is an implied share price: the fair value of NVIDIA equity under the chosen scenario, expressed in dollars per share. It answers the question, 'What should this stock trade at if our assumptions about earnings growth and market sentiment are correct?' The house number is the Base scenario result, currently $245–$278/share, which sits +9% to +24% above the $224 market price as of 7 August 2026.

It is the model's output implied_share_price, stated in $/share, computed as (fy2027_eps * (1 + fy2028_eps_growth / 100) * exit_pe) * (1 - lever_market_share_erosion * 0.165) from the argued parameters — not asserted in prose.

Market reference. 223.96 $/share as of 2026-08-07, from NASDAQ close via Bloomberg. Market data for reconciliation only — not house evidence.

8. The decision to publish

Gate outcome: overridden Claims refuted before publication: 3 — each was rewritten or removed Evidence grade: A — Evidence base is strong on financials (direct 10-Q/10-K filings) but relies on secondary sources for customer concentration details and carries one medium-confidence competitor claim that we cannot independently verify. Decay: Recheck against Q2 FY2027 results, 2026-08-26. Resolution date: 2026-08-26 Disclosure: CONTEXXT holds no position in NVDA and receives no compensation from any party named here. Research assisted by machine agents; every load-bearing claim is attributed and attacked before publication. Not published here: cost and token spend, retries and provider timeouts, editor identities, the system prompts, per-call model configuration, and the itemised list of discarded sources. None of it qualifies the research; all of it describes how we operate.