We asked Claude Fable 5.1 the same 8 questions about NVDA twice: once on its own, once with Finbly's verified evidence. Then code checked every figure in both sets of answers against the filings. The result is fewer hallucinations, current numbers, and 94% less spent on tokens.
Claude Fable 5.1 on both sides · 0 of Finbly's figures contradict the filings, 10 unverifiable · 94% less token spend · generated Sep 6, 2026 · data as of 2026-09-05
Short answer: yes — it dipped during the Blackwell ramp but is back in the low-to-mid 70s and guided to ~75%. (Figures through the Oct-2025 quarter, Q3 FY26; verify against the latest print.)
The trajectory (non-GAAP gross margin)
What drove the dip, and why it's reversing
What to watch
Bottom line: Gross margin bottomed at ~71% ex-one-offs in the April 2025 quarter and has recovered every quarter since, guided back to 75% — it's holding in the mid-70s, just not at the 78–79% Hopper peak, which was never the sustainable level.
Short answer: it's holding today, but management has told you it won't hold through next year. The last verified print was 74.9% gross margin (what's left of each revenue dollar after paying for the chips, memory and packaging that go into the product) in Q1 FY2027 [Q1 FY2027], comfortably above the 73% line that marks the bull case. On the August 26 call, though, NVIDIA guided that figure to slide to 74% ±50bps in Q3, bottom at 71–72% in Q4 FY2027, and recover only to 72–73% for all of FY2028 [earnings recap]. (The recap doesn't state the Q2 actual gross margin, so 74.9% is the last hard number.)
Why it's slipping — and this one is genuinely NVIDIA's own cost problem Unlike most input-cost shocks, memory hits NVIDIA's own cost of goods, because it buys HBM (high-bandwidth memory, the stacked DRAM bolted next to each GPU) from outside suppliers. The CFO said memory price increases "exceeded prior expectations and are headed even higher" [earnings recap, F07]. The backdrop: HBM4 runs ~$500/stack vs ~$300 for HBM3E [E], DRAM contract prices rose 93–98% quarter-on-quarter in 1Q26 [F], and all three HBM makers are sold out for 2026 [V]. NVIDIA is even cutting RTX 50-series gaming output to reallocate memory to data center [press]. Management has chosen to eat some of that cost rather than push all of it to customers.
Does this break the thesis? By the letter, yes — partly. The active trigger fires below 73% "absent a China charge" [watch], and this is a memory charge, not a China charge. The research note said a ~200bp slip would break the "pricing power" leg; 74.9% → 71–72% is roughly 300bps. Even the FY2028 recovery to 72–73% sits at or below the threshold. So the "margins hold in the mid-70s through the Rubin ramp" leg is impaired.
But the market has already voted that revenue matters more: shares rose 8.7% on 2.6x normal volume the day after that margin reset was disclosed [earnings recap, F02], because the same call guided FY2028 revenue growth of ~70% with demand described as supply-constrained (true demand ~100%) [earnings recap, F04]. Analysts followed — FY2028 consensus EPS rose 21% to $15.52 and FY2029 rose 35% in one week [estimate]. A 300bp margin haircut on revenue growing 70% is arithmetic the earnings line absorbs easily.
Context that softens the optics Full-year FY2026 gross margin was 71.1% [FY2026], but that was depressed by a one-time ~$4.5B China/H20 write-down; underlying was ~75% [F]. So a 71–72% trough is roughly where the reported figure already sat last year — the difference is that this time it's structural input cost, not a one-off. Net margin of 55.6% and ROIC of 66.8% [FY2026] leave enormous cushion; this is a question of how spectacular, not whether profitable.
What to watch
Bottom line: Gross margin is holding at 74.9% today, but by management's own guidance it will breach the 73% bull-case line in Q4 and stay near it through FY2028 on memory costs — the pricing-power leg is dented, and the stock's 8.7% post-earnings rally says investors are paying for 70% revenue growth, not for margin perfection.
Consensus target ~$170–175 (roughly 25–30% upside from mid-2025 levels ~$135), forward P/E ~30x on FY26 EPS. Figures below are as of my data cutoff (mid-2025) — this name moves fast, so verify against a live terminal.
Ratings and targets
Consensus estimates
Forward P/E
What to watch
Bottom line: Street is near-unanimously bullish with a ~$170–175 consensus target, and at ~30x FY26 / ~24x FY27 EPS the stock is priced as a GARP name, not a bubble — the whole thesis rests on estimate revisions staying positive.
Analysts see ~50% upside: average price target $345 (median $322.50) against a $230.36 stock, and a forward P/E of roughly 25x — the cheapest NVDA has looked on next-year earnings in a decade. [estimate] [state]
Price targets
Forward P/E (price ÷ next year's expected earnings)
Estimates are moving up hard, not down
Two things to hold alongside the optimism
Bottom line: The Street is unanimous and getting more bullish — ~$345 average target, ~25x forward, estimates rising 20–90% out-year — but the stock's low multiple isn't a gift; it's the market pricing in real doubt about whether those out-year numbers survive the financing and power constraints downstream.
NVIDIA Q3 FY2026 (quarter ended Oct 26, 2025; reported Nov 19, 2025) — a clean beat-and-raise with acceleration, not deceleration.
Headline numbers
Segments
Guide / forward color
What it told you
Note: this is the latest quarter in my data. If Q4 FY26 (due late Feb 2026) has already been reported, check it against the $65B guide.
Bottom line: $57B revenue (+62%), $51B Data Center, $1.30 EPS, and a $65B Q4 guide well above Street — fundamentals accelerated; the stock's issue is multiple/sentiment, not the numbers.
The quarter was a clean beat with a bigger guide, and the stock jumped 8.7% on it — but management also cut the margin outlook, which is the one thing worth dwelling on.
The print (FY2027 Q2, reported Aug 26)
The guide (the real driver of the reaction)
The soft spot: gross margin
Also raised on the call
Where that leaves valuation
Bottom line: Demand and growth are not the debate — a $96B quarter, a $108B guide, and ~70% FY28 growth settled that; the stock's 8.7% pop and 20–35% estimate upgrades reflect it. The quarter's one new negative is real: memory inflation is pushing gross margin toward 71–72% by Q4, below the 73% line that defines "pricing power intact," and that plus the $50B of customer financing is where the risk now lives.
Short answer: roughly 28–30x forward earnings — expensive on an absolute basis, not expensive relative to its growth. (Figures below are from my most recent data; sanity-check against today's price.)
The multiples
Versus history and peers
What the price assumes
Where it gets demanding
Bottom line: At ~28–30x forward with 40%+ growth, NVDA is priced for continued execution, not perfection — it's cheap on PEG and demanding only if you think 2026 hyperscaler capex is the peak.
Short answer: demanding in absolute terms, but not on NVIDIA's own history or relative to its growth rate — the price already bakes in ~30% annual cash-flow growth for a decade, yet it's the cheapest the stock has been on forward earnings in ten years.
What the price embeds At $230.36 [state], a reverse DCF — working backwards from the share price to find the growth rate needed to justify it — says NVIDIA must grow free cash flow about 29.4% a year for ten years [reverse-DCF], using a ~14% discount rate (the return investors demand for the risk) [valuation agent]. That's a high bar: starting from $96.7B of FY2026 free cash flow [F], a decade of ~29% compounding is a very large number. The trigger level is 32% [watch]; at 29.4% you're inside the line, but with little slack — a further run-up in price without a matching rise in cash flow pushes the market's demand past what the watch list considers "too much."
The multiples
Against its own history: cheap NVIDIA's forward P/E has averaged ~63x over five years and ranged ~38–120x since FY2022 [F]; today's ~21–25x is the lowest in a decade [F]. The stock has spent the past year letting earnings grow into the price rather than re-rating — roughly flat YTD while AMD was up 130%+ [F]. In plain terms, at ~25x forward on 65–85% growth, the market is pricing either a peak-cycle earnings year or a structurally lower long-run margin than today's 74.9% [Q1 FY2027]. The bull case doesn't need the multiple to expand — it just needs FY2028 earnings not to fall below FY2027's [F]. Consensus currently has FY2028 EPS at $15.52, up from $9.27 in FY2027 [estimate], and revisions ran 10 up / 0 down in the last week, with the average target lifted to $345 (~50% upside) [estimate].
Against peers: very expensive Value NVIDIA at the median semiconductor P/S and you get roughly −67.5% [comps] — that trigger has FIRED (threshold −65%), meaning the premium to peers is stretching further [watch]. The quality justifies a premium — ROIC 66.8% [state] against a ~14% cost of capital is a ~53-point spread [valuation agent] — but this is the floor if the market ever decides it's "just a chip company."
Why the market won't pay more The doubt isn't demand; it's whether the buyers can keep paying. Hyperscaler free cash flow is heading toward zero in 2026 with capex debt-funded [E]; ~$16B of NVIDIA's own GAAP quarterly profit was unrealized gains on customers it has invested in [F]; and the reported ~$250B OpenAI deal raised "circular financing" fears, knocking the stock ~4% on July 27 [watch]. Insiders sold a net 2.8M shares with zero open-market buys in six months [state]. Those are the reasons a 25x forward multiple coexists with 29% embedded growth.
Bottom line: The valuation demands ~29% free-cash-flow growth for a decade — a real bar — but it sits at a decade-low forward multiple on 65–85% growth, so the stock is priced for the growth to slow sharply, not to continue; the risk isn't the multiple, it's whether debt-strained customers keep funding the buildout.
Short answer: five risks matter, and the two that could actually break the story are hyperscaler capex digestion and China/export policy. Everything else is a multiple-compression risk, not a thesis-breaker.
1. Customer concentration + capex cycle (biggest) Data center is ~88% of revenue, and the four largest hyperscalers are roughly 40–50% of that. Combined big-4 capex is running ~$300B+ annualized. The risk isn't that AI spending stops — it's a pause: a single quarter of "digestion" guidance from MSFT/META/GOOG/AMZN would flow straight into NVDA's sequential growth. Watch capex commentary on hyperscaler calls more than NVDA's own guide.
2. China / export controls China was ~20–25% of data center revenue before the 2022–23 restrictions; it's now mid-single to low-double digits and NVDA took a ~$4.5–5.5B write-down on H20 inventory when licenses were pulled. The remaining exposure is asymmetric: further tightening (or Beijing steering buyers to Huawei Ascend) removes revenue, while any loosening is upside not in most models.
3. Custom silicon and AMD Google TPU, Amazon Trainium, Meta MTIA, Microsoft Maia — Broadcom's custom-ASIC AI business alone is ~$12B+/yr. NVDA still holds ~80–90% of merchant accelerator share, but the inference workload (where ASICs compete best) is growing faster than training. A shift from ~85% to ~70% share over three years would be enough to cap the multiple even with a growing market.
4. Gross margin normalization GM peaked ~78–79% (Q1 FY25) and has been guided into the low-to-mid 70s through the Blackwell ramp. Company says mid-70s is the new steady state; consensus largely believes it. If rack-scale systems (GB200/NVL72) and HBM cost inflation push it toward 70%, EPS growth decouples from revenue growth.
5. Supply chain / Taiwan Effectively single-sourced on TSMC (CoWoS packaging is the binding constraint) with HBM from SK Hynix/Micron/Samsung. Blackwell's initial mask-set flaw cost roughly a quarter. Any Taiwan disruption is unhedgeable — tail risk, but a fat one.
Second-tier (watch, don't lose sleep):
Bottom line: NVDA's risk is a demand-timing risk concentrated in ~5 buyers and one geopolitical relationship — not a competitive-displacement risk yet — so the stock is far more sensitive to a hyperscaler capex "pause" headline than to anything AMD or Broadcom ships.
The five biggest live risks, in order: (1) the financing behind the AI buildout, (2) China, (3) competition for the inference tier, (4) supply/memory, and (5) a stretched valuation with insiders selling. Notably, none of them is "demand" — Q1 FY2027 revenue was $81.6B (+85% YoY) with a 74.9% gross margin, and Q2 was guided to $91.0B [Q1 FY2027] [F]. The risks all sit one layer downstream of NVDA's own P&L.
1. Financing and "circular" demand — the highest-signal risk
2. China — optionality removed, antitrust hanging over Mellanox
3. Competition / share loss in inference
4. Supply chain and single-source dependence
5. Valuation and insider behavior
Bottom line: NVDA's own numbers are pristine; the live risks are whether its customers can keep paying for the buildout without NVDA financing them, whether China turns from lost upside into active penalty, and whether the market keeps refusing to pay for growth it doesn't dispute.
Scorecard (through the Q3 FY26 call, Nov 2025): ~7 promises delivered, 2 revised, 1 outright miss. Revenue guidance has been beaten 7 of the last 7 quarters by $1.5–3B; the only real broken promise is China. Verify anything after Nov 2025 — I don't have the Q4 FY26 (Feb 2026) print.
Delivered
Revised
Missed
Bottom line: Management has under-promised and over-delivered on product, revenue and margins; the one miss is China, and the one big unproven promise is the $500B booking figure — track that against 2026 data center revenue.
Short version: NVIDIA's track record on product and revenue promises is close to spotless; its track record on cost promises is not. Of 31 commitments made across the last six earnings calls (May 2025–Aug 2026), 17 have resolved and 14 are still open [transcripts]. Every product-timing and revenue target in the resolved set was delivered — several were beaten outright. Every miss is the same miss, made four times: operating-expense growth guided too low and then revised up.
Delivered (7 of the resolved set shown)
On track (3 open items with evidence)
Revised / withdrawn (4 — all OpEx)
Nothing was outright abandoned, but the pattern matters: costs keep running ahead of plan, twice in a row, in the same direction. With a 55.6% net margin [FY2026] this hasn't dented profits, but it's a small crack in the "operating leverage" story.
The new promises to hold them to (from the Aug 26 call)
Why the market forgave the margin reset: the print beat on EPS ($2.22 vs $2.09) and revenue ($96.2B vs $92.3B), the seventh straight double beat, and the stock rose 8.7% on 2.6x normal volume [earnings recap]. Consensus FY2028 EPS jumped 21% to $15.52 in the week after [estimate]. At $230 the stock trades on 24.8x forward earnings [state] — revenue growth is doing the heavy lifting, not margin expansion, which is why the 70% FY28 revenue promise is now the single most important commitment on the books.
Bottom line: Management has delivered every product and revenue target it set and beaten the big ones; the only misses are four consecutive upward OpEx revisions plus a pre-emptive margin reset to 71–73% — a cost-discipline story, not a demand story, and the stock has judged it accordingly.
Framing: I'm comparing the Q3 FY26 10-Q (quarter ended Oct 26, 2025, filed Nov 2025) against the Q2 FY26 10-Q (quarter ended Jul 27, 2025). If a newer 10-Q has been filed since my data, verify these against the current document. Figures are from memory of the filings and are approximate where noted.
Headline numbers that moved
New/changed disclosures that matter most
Supply commitments and inventory (the demand signal). Inventory rose to ~$19.8B from ~$15.0B; total inventory purchase obligations plus long-term supply/capacity commitments rose to ~$50.3B from ~$44B. Together with the "$500B of Blackwell + Rubin visibility through CY2026" language, this is management leaning into supply. It's bullish on demand but raises the write-down tail risk if hyperscaler capex slips (recall the $4.5B H20 charge in Q1).
Receivables build. A/R jumped to ~$33.4B from ~$23.1B, DSO drifting into the low-50s. Operating cash flow (~$23.8B) lagged net income. Not alarming given quarter-end shipment timing to a handful of large customers, but it's the line bears are watching for signs of stretched customer payment terms.
Customer concentration and "circular" financing. Two direct customers were ~22% and ~15% of revenue (vs ~23%/~16% in Q2), with additional indirect customers estimated >10%. New in Q3: the non-binding LOI to invest up to $100B in OpenAI (first $10B tied to first-gigawatt deployment) and the $5B Intel equity purchase (pending regulatory approval at filing). The 10-Q now explicitly frames these as strategic investments rather than customer financing, but the concentration + vendor-investment combination is the core thesis risk to monitor.
China. No H20 China revenue for a second straight quarter; the filing notes Chinese government actions discouraging NVIDIA purchases and that the reported 15% revenue-share arrangement with the U.S. government had not been formalized. China is effectively a free option in the model now, not a base-case contributor.
Capital return. ~$12.5B repurchased in Q3 (vs ~$9.7B in Q2), ~$62B authorization remaining after the $60B August add. Buybacks are now roughly offsetting SBC dilution and then some.
Risk-factor additions. Expanded language on Section 232 semiconductor tariffs, AI regulatory scrutiny (U.S. and EU), and power/data-center infrastructure constraints as gating factors for customer deployments.
Bottom line: The Q3 10-Q shows accelerating revenue and margin recovery, but the thesis-relevant changes are the ~$50B supply commitment book, a ~45% QoQ jump in receivables, and newly disclosed equity investments in customers/partners—demand visibility is stronger than ever, and so is the concentration risk if that demand proves circular.
Short answer: eleven material changes, and the one that matters most is a brand-new $36 billion of commitments NVDA has made to buy back cloud capacity from its own customers. That single disclosure turns the "circular financing" worry — Nvidia's money funding Nvidia's demand — from a rumor into a line item [SEC filings][watch]. Everything else in the filing is secondary to it, though China and the Rubin ramp both deserve a look.
What changed (10-Q filed 2026-08-26 vs 2026-05-20) 28 candidate changes reviewed, 16 mechanical ones suppressed, 11 flagged as material [SEC filings]. Ranked by thesis impact:
1. New "AI cloud partner" business model — $36B of capacity commitments (highest impact) NVDA now sells data-center systems to select AI clouds (the "neoclouds" — smaller cloud providers built around renting out GPUs) and in return commits to buy cloud service from them, typically for six years. Commitments totaled $36B as of July 26, 2026, and shrink as third parties use the capacity. The clouds can unilaterally stop providing capacity to NVDA and sell it elsewhere at better rates [SEC filings]. Why this is the crux: it's vendor financing in all but name — NVDA is underwriting demand for its own chips. It's also exactly what the open watch item was waiting for after the reported ~$250B OpenAI deal knocked the stock ~4% on 2026-07-27 [watch]. The escalation condition on that trigger was "confirmed large vendor financing at aggressive terms" — six-year commitments the partner can walk away from look like aggressive terms. It also explains why the CoreWeave credit-default-swap trigger (the cost of insuring a neocloud's debt, a stress gauge) is still marked FIRED at ~452bps after an ~881bps peak [watch]: if those partners wobble, NVDA is now on the hook.
2. MOUs to mobilize $500B+ of third-party AI infrastructure capital In August 2026 NVDA signed memorandums of understanding (non-binding preliminary agreements) with large capital providers to build financing platforms for AI buildout, independently underwritten. NVDA "may provide limited residual-value support" — i.e., it may guarantee what the hardware is worth at the end of a project [SEC filings]. Bull read: it widens the pool of buyers who can afford Rubin-class systems. Bear read: same mechanism as #1, scaled up — NVDA absorbing tail risk on assets it sells. There's no figure for how much residual support might be extended, so the exposure isn't quantifiable yet.
3. China: H20 sales blocked by Beijing despite US licenses The filing confirms the PRC prevented H20 sales even after Washington granted licenses, so the inventory couldn't be sold — reversing the prior hope of a licensed China revenue stream [SEC filings]. The China antitrust language was also broadened into generalized export-control risk, including the possibility of retaliation hitting NVDA's supply chain [SEC filings]. This effectively closes the "China thaw" optionality the base case already excluded [watch]; it's a loss of upside rather than a new hit to numbers, since FY2026's 71.1% gross margin already absorbed a ~$4.5B H20 write-down [curated memory].
4. Vera Rubin production shipments began (Q3 FY27), supply still constrained The next-gen architecture is shipping alongside Blackwell, with NVDA disclosing supply constraints, large multi-year supply commitments, and risks of yield issues and inventory provisions [SEC filings]. Read this together with guidance: the quarter beat handily (revenue $96.22B vs $92.27B, EPS $2.22 vs $2.09, ~70% FY28 growth guided) but gross margin is guided down to 71-72% in Q4 FY27 on memory-cost inflation, from 74.9% in Q1 FY27 [company analyst digest][Q1 FY2027]. The gross-margin trigger sits at 73% — a print below that "absent a China charge" breaks the margins-hold-through-Rubin leg of the bull case [watch]. So the 10-Q disclosure and the guidance point at the same fault line.
Why this matters against the price At $230.36 the stock is 45.1x trailing earnings but only 24.8x forward [state], and the reverse DCF (working backward from the price to the growth it requires) says the market is baking in ~29.4% annual free-cash-flow growth for ten years [reverse-DCF]. That's why the bull/bear debate is scored bear-leading: the price already assumes near-flawless execution [debate agent]. Changes #1 and #2 add a new way to stumble — demand that's partly financed by the seller — and #4 trims the margin cushion. Offsetting that, the market voted the other way: shares rose 8.7% on the print and estimates moved up across the board (FY2028 EPS +21% to $15.52, average target $345 = ~50% upside) [company analyst digest][estimate][state].
Bottom line: The filing's headline change is $36B of self-underwritten demand plus a $500B financing framework with possible residual-value guarantees — the circular-financing risk is now real and disclosed — while China optionality is gone and the margin bar has dropped to 71-72%; the beat and revisions keep the bull case alive, but the thesis now hinges on whether those cloud partners stay solvent and Q4 margins hold above 73%.
Short version: NVDA's single point of failure is TSMC (100% of leading-edge wafers and the CoWoS advanced packaging that gates Blackwell/Rubin output), followed by HBM (SK Hynix ~55–60% share of NVDA's supply) and a customer base where the top 4–5 hyperscalers are ~50% of Data Center revenue. Note: figures below run through the Oct-2025 quarter (Q3 FY26); verify anything after that.
Upstream – where the real choke points are
Midstream – rack assembly
Foxconn is the dominant GB200/GB300 NVL72 integrator, then Quanta, Wistron/Wiwynn, then Dell and Supermicro as branded OEMs. GB200 rack yield issues (backplane, liquid cooling) delayed volume by ~2 quarters into 1H 2025; GB300 was fixed by sharing the GB200 form factor. Supermicro's audit crisis (EY resigned Oct 2024, 10-K filed Feb 2025) shifted some share to Dell. Liquid cooling (Vertiv, Delta) is a new dependency – NVL72 can't run air-cooled.
Downstream – customer concentration is worse than the 10-K headline
Offsets
Backlog visibility: management cited ~$500B of Blackwell + Rubin orders through CY2026. Customer diversification is improving at the margin (sovereigns, enterprise, neoclouds), and NVDA's ~$60B+ in supply prepayments/purchase commitments locks capacity ahead of AMD and custom-ASIC programs (Broadcom/Google TPU, Trainium) that compete for the same CoWoS and HBM.
Bottom line: The supply side is a TSMC-CoWoS-plus-Hynix duopoly problem that's easing but not solved until 2026; the demand side has shifted from "too concentrated in five hyperscalers" to "increasingly financed by NVDA itself," and China is a zero – so the stock's supply-chain risk is now less about missing parts and more about Taiwan and customer quality.
The short answer: NVIDIA's chain is concentrated at both ends — a handful of Asian suppliers feed it, and a handful of hyperscalers and AI labs buy from it — and the thing that changed most recently is that supplier tightness has started showing up in NVIDIA's own margins, while customer risk has shifted from "will they buy" to "how are they paying."
Supply side: three chokepoints
Further downstream, the binding constraint through 2027 is power, not silicon — transformer lead times ~128 weeks [E] and grid interconnection queues of 4–7 years [F]; only ~5 GW of ~16 GW of announced 2026 US capacity is under construction [E].
Customer side: concentration plus circularity
What changed in the last few weeks
Bottom line: Supply concentration (TSMC, HBM, CoWoS) is a cost and cadence risk that just cost NVIDIA ~3 points of margin guidance; customer concentration is a financing risk sitting one layer downstream — and with the stock at 45.1× trailing / 24.8× forward and 29% growth already embedded, a coordinated hyperscaler capex cut or a second quarter of memory-driven margin slippage is what actually breaks the thesis.
5 SEC filings, 9 earnings calls and 134 headlines, plus 1,129 verified financial facts across 5 fiscal years. About 288 of those pages are the evidence behind a single answer. You never paste any of it.
| Document | Filed | Pages | Tokens |
|---|---|---|---|
| filing10-K · period ended 2026-01-25 | 2026-02-25 | 114 | 120.7k |
| filing10-Q · period ended 2026-07-26 | 2026-08-26 | 51 | 66.1k |
| filing10-Q · period ended 2026-04-26 | 2026-05-20 | 46 | 57.6k |
| transcript4 earnings calls · Q3 FY2026, Q4 FY2026, Q1 FY2027, Q2 FY2027 | 2026-08-26 | 73 | 62.5k |
| news123 headlines, last 30 days | 2026-09-06 | 4 | 6.2k |
| Question | Chatbot | Finbly Ask | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Cited | Match | Wrong | Unverifiable | Cost | Cited | Match | Wrong | Unverifiable | Cost | |
| 01Is the gross margin holding? | 0 | 19/22 | 0 | 3 | $0.076 | 11 | 36/37 | 0 | 1 | $0.22 |
| 02What are analysts expecting — targets and forward P/E? | 0 | 6/32 | 4 | 22 | $0.066 | 18 | 32/38 | 0 | 6 | $0.15 |
| 03What did the most recent quarter show? | 0 | 21/46 | 0 | 25 | $0.084 | 11 | 54/55 | 0 | 1 | $0.16 |
| 04How demanding is the valuation right now? | 0 | 10/34 | 2 | 22 | $0.080 | 26 | 41/42 | 0 | 1 | $0.16 |
| 05What are the biggest live risk factors? | 0 | 5/16 | 1 | 10 | $0.10 | 21 | 30/31 | 0 | 1 | $0.18 |
| 06Which promises has management made on recent earnings calls, and which have they delivered, revised or missed? | 0 | 9/38 | 0 | 29 | $0.18 | 4 | 31/31 | 0 | 0 | $0.22 |
| 07What changed in the latest 10-Q versus the prior one, and which new disclosures matter most for the thesis? | 0 | 10/42 | 0 | 32 | $0.19 | 10 | 31/31 | 0 | 0 | $0.20 |
| 08Trace the supply chain: where are the concentration risks, and what has changed recently with key suppliers and customers? | 0 | 13/28 | 0 | 15 | $0.21 | 18 | 32/32 | 0 | 0 | $0.21 |
| Model | Chatbot | Finbly Ask | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Cited | Match | Wrong | Unverifiable | Cost | Cited | Match | Wrong | Unverifiable | Cost | |
| 8 qClaude Fable 5.1 | 0 | 93/258 | 7 | 158 | $0.98 | 119 | 287/297 | 0 | 10 | $1.51 |
| 8 qGPT-5.6 Sol | 0 | 40/104 | 6 | 58 | $0.18 | 72 | 149/163 | 0 | 14 | $0.44 |
| 8 qClaude Sonnet 5 | 0 | 34/83 | 11 | 38 | $0.12 | 53 | 152/155 | 0 | 3 | $0.39 |
| 8 qK2.6 | 0 | 59/151 | 6 | 86 | $0.015 | 53 | 147/150 | 0 | 3 | $0.050 |
Code pulls every number out of both answers. A figure whose surrounding words name a metric Finbly tracks is compared with the verified value and passes within 3%. Other figures are looked up in the evidence the grounded model received. A figure tied to an earlier quarter is marked stale. No model judges the result. The chatbot side ran the same model with the same tone instructions and no data.