Every element here was extracted from a published report, attacked by our adversary, approved by an editor and placed in the article. Each carries its provenance and a link back to the report and its verification annex. Quote them, embed them — the attribution travels with the claim.
At 253p and approximately £555m market capitalisation, the equity embeds 247p per share of SpaceX expectation over a 6.2p non-SpaceX floor, and would require £324m of FY2027 revenue at the house margin to justify today's price.
CONTEXXT · house analysis · internal valuation model · 2026 · e31
Filtronic PLC reported operating profit of £4.0m for the fiscal year ending May 2026, a decline of 70% from the prior year's £13.3m.
AskTraders.com · financial news outlet · Filtronic Shares Tumble Despite Record Order Book as Investment Squeezes Profits · 2026-08-04 · e7
Filtronic held net cash of £10.5m at FY2026 year-end [e14b] — about 2% of market capitalisation at 253p. The £11.3m figure was refuted at verification.
Filtronic plc · company FY2026 results statement · Filtronic FY2026 preliminary results, net cash at year-end · 2026-06-23 · e14b
Edison Investment Research forecasts Filtronic PLC will generate adjusted EBITDA of £11.8m in the fiscal year ending May 2027, implying an EBITDA margin of 20% on consensus revenue of £59.6m, which remains 1,000 basis points below the 30% margin achieved in FY2025 when EBITDA was £17.0m on revenue of £56.3m.
Edison Investment Research · equity research house · research note · 2026 · e11
Filtronic PLC reported adjusted EBITDA of £11.3m for the fiscal year ending May 2026 [e16], a decline of £5.7m or 34% from £17.0m in FY2025, while revenue remained flat at £55.5m across both periods.
Filtronic PLC · issuer · company disclosure · 2026 · e16
At 253p the market pays £555m for a business with £10.5m of net cash [e30] and no disclosed SpaceX contract term — 247p a share of permanence, bought without documentation.
Provenance not recorded
Flat revenue and a 70% operating-profit decline: the margin collapsed, the top line did not. Whether that was investment or repricing is the whole question, and only the FY2027 print answers it.
Provenance not recorded
Reducing SpaceX from 83% to 70% of revenue means a 90% jump in everything else within one year, on the strength of a June 2026 win of undisclosed size.
Provenance not recorded
£10.5m of net cash is 2% of the market capitalisation: a cushion, not a floor. The FY2027 print in August 2027 is the only thing that can settle the permanence bet.
Provenance not recorded
Two unnamed customers together accounted for approximately 13% of NVIDIA's total revenue in Q1 fiscal 2027, the quarter ended 26 April 2026, down from almost 40% concentration in Q2 fiscal 2026 ended 27 July 2025—a 27 percentage-point reduction in customer concentration over three quarters.
NVIDIA Corporation · 10-Q filing disclosure · Form 10-Q for quarter ended 26 April 2026 · 2026-04-26 · e5
NVIDIA reported GAAP gross margin of 74.9% / 75.0% non-GAAP [e22] for Q1 fiscal 2027 ended 26 April 2026, roughly 300 basis points below the 78.0% data center gross margin for the full fiscal 2026 year ended 25 January 2026 — corrected under Editor memo 2026-08-07
NVIDIA Corporation · 10-Q filing financial results · Form 10-Q for quarter ended 26 April 2026 · 2026-04-26 · e22
The 5.9 million developers using CUDA as of FY2026 [e3] is the figure most often cited as proof of moat, but it is the wrong metric: developer count is a lagging indicator — sticky, yes, but silent on pricing power in real time. Gross margin is not. NVIDIA reported 74.9% GAAP / 75.0% non-GAAP [e22] gross margin in Q1 FY2027 [e18], roughly three points below the 78% recorded for the full FY2026 [e21] — corrected under Editor memo 2026-08-07. A margin still above every falsifiability line is the live signal that hyperscalers have not yet forced pricing concessions and that AMD competition has not compressed realised ASPs. The Q2 FY2027 print on 26 August 2026 will settle the question. A gross margin below 70% would confirm that customer concentration — approximately 13% of total revenue in two unnamed customers in Q1 FY2027 [e5], and approximately 22% in one direct customer across FY2026 [e12] — has translated into negotiating leverage, and would validate the antithesis that dual-sourcing to AMD and custom ASICs is forcing NVIDIA to discount. A hold above 75% would show the moat has migrated to the hardware-software stack — NVLink, Tensor Cores, the Mellanox networking fabric acquired for $6.9 billion in March 2019 [e6]. The mechanism is not proven. Two unnamed customers accounted for almost 40% of Q2 FY2026 revenue [e9], yet gross margin held at 78% for the full FY2026 [e21]: either hyperscalers have not yet exercised their leverage, or NVIDIA's differentiation still resists it. The Cerebras CEO's claim that CUDA has lost approximately 70% of frontier AI training share [e7] is directionally troubling, but the margin data through Q1 FY2027 does not corroborate it. The second indicator is growth. Data centre revenue grew 92% year-over-year in Q1 FY2027, rising from $39.1bn [e24] to $75.2bn [e23] — corrected under Editor memo 2026-08-07 — and the comparison already flatters: Q1 FY2026 marked the early ramp of Hopper deployments, so the delta reflects both volume growth and the architectural transition from Ampere. Nearly doubling revenue in twelve months cannot persist without NVIDIA capturing every incremental dollar of hyperscaler AI capital expenditure. We treat 60% as the threshold because it implies NVIDIA still captures the majority of incremental spend, while acknowledging that the installed base is now large enough that doubling it annually would require hyperscaler AI budgets to grow at rates inconsistent with their own disclosed capital plans. Growth above that level extends the moat thesis; growth of 40-50% would confirm that dual-sourcing has moved from pilot programmes to production scale. Deceleration would show up first where the CUDA switching cost is lowest, in inference. Any one of the top five accounts shifting 20-30% of new orders to AMD or internal ASICs would be visible in the quarterly growth rate: the disclosed two-customer share fell to approximately 13% in Q1 FY2027 [e5] from almost 40% in Q2 FY2026 [e9, e14, e20], while the absolute revenue contribution of the top accounts grew. Both indicators land in the same print.
Provenance not recorded
NVIDIA's moat has shifted from CUDA-the-API to hardware-software co-design — and the question is whether the $320–350 the thesis implies is already in the price. Our answer after verification: the house Base says $261 against a $224 market, and the market is priced closer to the antithesis.
nick@contexxt.com · editor · Editor memo 2026-08-07 · 2026-08-07 · e22
We no longer believe CUDA-the-API is the defensible asset. The 5.9 million developers using CUDA and NVIDIA's other software tools as of FY2026 [e3] represent installed base, not switching cost. Cerebras CEO Andrew Feldman claimed in the week ending 1 June 2026 that CUDA has lost approximately 70% of frontier AI training share [e7] — a figure we cannot verify independently, but one that aligns with the hyperscaler shift to custom ASICs for training workloads. The new moat is vertical integration at the cluster level: NVLink interconnects, Tensor Cores optimised for transformer architectures, and the NeMo framework that co-designs model checkpointing with NVLink topology. NVIDIA paid $6.9 billion for Mellanox in March 2019 [e6], a deal initially framed as defensive to secure the InfiniBand stack. The switching cost is no longer rewriting CUDA kernels — it is replacing an entire networking fabric and re-sharding training runs across a different interconnect topology. We cannot quantify that cost with the evidence in hand. The $15-25 million per 10,000-GPU cluster figure in the section brief does not appear in any filing or third-party source we retrieved, and we will not defend a number we cannot cite; we flag it as our own model output rather than observed data. The $6.9 billion Mellanox price, amortised across the installed base, bounds what NVIDIA spent to build the integration — not what a hyperscaler would spend to route around it. The real test is whether a hyperscaler building a custom ASIC can forgo NVLink and InfiniBand entirely, or must replicate the fabric at comparable cost. We have no public bill-of-materials for a Google TPU pod or an Amazon Trainium cluster, and no filing that breaks out capital or operating expense attributable to networking versus silicon. The margin evidence — argued under The number to watch [e22][e21] — says pricing power has not cracked. We believe the moat exists; we cannot yet price it.
Provenance not recorded
NVIDIA's customer base remains alarmingly narrow, even as the disclosed concentration has fallen: two unnamed customers accounted for approximately 13% of total revenue in Q1 FY2027 [e5], down from almost 40% in Q2 FY2026 [e9, e14, e20]. One direct customer represented approximately 22% of total revenue across the full FY2026 fiscal year [e12]. What has grown is the dollar dependence — the same handful of hyperscalers writes materially larger cheques on a revenue base that nearly doubled — and it is that absolute dependence, not a worsening ratio, that gives buyers their negotiating leverage. We cannot confirm whether the same accounts appear across these periods; NVIDIA does not name them. If five hyperscalers still represent 45-55% of revenue — a figure we infer from the disclosed concentrations but cannot cite directly — pricing power rests with the buy side regardless of switching costs. This is not a diversification story. The 5.9 million developers using CUDA and NVIDIA's other software tools as of FY2026 [e3] represent ecosystem breadth, not revenue breadth. The actual revenue flows through a bottleneck of five or fewer accounts, and those accounts are building their own silicon, dual-sourcing to AMD, and sitting across the table from NVIDIA with full visibility into margin structure. A supplier still earning the margin argued under The number to watch [e22][e18] is a supplier with room to concede price. The risk is not that hyperscalers walk away — they cannot, not in volume, not yet — but that they extract concessions in exchange for continued orders. Volume rebates, co-engineering commitments, or simply the implicit threat of shifting incremental capacity to Instinct or Trainium: any of these can compress realised pricing without showing up as a list-price cut. Cerebras CEO Andrew Feldman claimed in the week ending 1 June 2026 that CUDA had lost approximately 70% of frontier AI training share [e7]. We treat this as a directional signal rather than a settled fact: Cerebras has every incentive to talk up custom silicon, the claim carries no denominator, and the company offered no customer names or workload breakdowns. The direction of travel is consistent with frontier labs splitting training runs across multiple architectures to derisk single-vendor lock-in. We separate training from inference because the switching costs differ by an order of magnitude. Training workloads are batch-oriented, run for weeks or months, and can tolerate bespoke toolchains; inference is latency-sensitive, customer-facing, and requires the optimised libraries that only CUDA and, to a lesser extent, AMD ROCm provide today. NVIDIA's data centre revenue growth through Q1 FY2027 — argued in full under The number to watch [e24][e23] — shows no sign yet that custom ASICs have inflicted volume loss at the consolidated level. Inference is where hyperscalers have the strongest economic motive to dual-source: the workloads are predictable, the models are frozen, and the margin on serving a query is thinner than on selling a training cluster. If AMD or custom ASICs capture 30-40% of incremental inference deployments by the end of calendar 2027 — the scenario the antithesis posits — data centre growth would decelerate sharply even if the moat on training remains intact. Until then, the Cerebras claim is a useful warning flag, not a falsified thesis.
Provenance not recorded
NVIDIA's data center revenue grew 92% year-over-year to $75.2bn [e23] USD in Q1 fiscal 2027 ended 26 April 2026, up from $39.1bn [e24] USD in Q1 fiscal 2026 ended 25 April 2025, an absolute increase of $36.1 billion USD in twelve months — corrected under Editor memo 2026-08-07.
NVIDIA Corporation · 10-Q filing segment revenue disclosure · Form 10-Q for quarter ended 26 April 2026 · 2026-04-26 · e22
Over 5.9 million developers worldwide used CUDA and NVIDIA's other software tools as of fiscal 2026 ended 25 January 2026, according to the company's annual filing, though NVIDIA does not disclose the methodology for counting active developers or distinguish between hobbyist and enterprise users.
NVIDIA Corporation · 10-K filing ecosystem metric · Form 10-K for fiscal year ended 25 January 2026 · 2026-01-25 · e3
The thesis and antithesis converge on one empirical question: what margin trajectory is already in the price? The margin evidence — argued under The number to watch [e22][e21] — says concessions have not yet been extracted despite the concentration argued under The buyers [e9, e14, e20][e12] — corrected under Editor memo 2026-08-07. The Q2 FY2027 print on 26 August 2026 will settle whether that concentration is rising — two unnamed customers accounted for approximately 13% of total revenue in Q1 FY2027 [e5], a sharp drop that either signals diversification or reflects a disclosure quirk in how NVIDIA aggregates cloud-service-provider resellers. If gross margin falls below 70%, the antithesis is correct: dual-sourcing to AMD and custom ASICs is forcing price cuts that no amount of vertical integration can offset. If margin holds above 75% while revenue growth exceeds 60% year-over-year, the thesis holds and the current multiple is underpricing the duration of the moat through fiscal 2028. $223.96 (2026-08-07) maps onto roughly +20% FY2028 EPS growth at a 20–22x multiple — the market is already pricing meaningful moat decay and multiple compression, and its position sits closer to the antithesis than the thesis. The derivation of the house $320–350 band, and the memo passages quoted in support of it, are relocated to the methodology annex, marked relocated by editor. Asymmetry, priced at the stamped reference $223.96 (2026-08-07): Thesis $322.40 (+44%) against Antithesis $177.77 (−21%) — 2.1:1 in favour of the upside leg. The framework favours the long side if margin persistence is assigned better-than-coin-flip odds; it says nothing about what those odds are.
Provenance not recorded
Every rule here was learned the expensive way: a defect on a live run, an editor's ruling, and a change to the code. Each entry names the run, the defect, the invariant it became and the check that now enforces it. Where nothing enforces it yet, it says so.