NVIDIA — Record Q3 Cools, DSI Re-Expands to 221 Days — Research Note | Silicon Nexus
Research Notes· May 16, 2026· NVDA· 9 min read
NVIDIA — Record Q3 Cools, DSI Re-Expands to 221 Days
Quarterly revenue stepped down 33% from a $91B peak as inventory days widened sharply, against accelerating Korean semi exports.
At a Glance
NVIDIA closed its most recent reported period, 2026-Q4, with revenue of $60,922 million, operating income of $32,972 million (54.1% operating margin), and net income of $29,760 million. Diluted EPS came in at $1.19. The print followed an exceptional 2026-Q3 in which revenue reached $91,166 million and operating income $57,419 million; in other words, the sequential top line stepped down 33.2% from the peak quarter even as absolute dollar profitability remained extraordinary by any historical standard. Days sales of inventory (DSI) widened to 221 in 2026-Q4 from 125 in the prior quarter, a swing that warrants attention. The share price moved from $186.23 on 2026-01-16 to $225.32 on 2026-05-15, a 21.0% gain over 83 trading days. The most recent SEC filings have been three low-impact officer/director 8-Ks (the latest on 2026-05-08), the FY2026 10-K (2026-02-25), and the FY2026-Q4 earnings release on the same date. No medium- or high-relevance external news items where NVDA is the affected ticker appear in the trailing 90-day filter.[1][2]
Business & Position
NVIDIA designs GPUs, AI accelerators, and networking silicon for data center, gaming, and automotive applications. Its position in the semiconductor stack is that of a fabless architecture house: it depends on TSMC for advanced-node wafer supply (and, critically, for CoWoS-class advanced packaging), on SK Hynix, Samsung Electronics, and Micron for high-bandwidth memory (HBM), and on assembly and test specialists including ASE Technology for backend. Downstream, revenue concentration sits with a cluster of hyperscale customers — Microsoft, Meta Platforms, Alphabet, Amazon, and Oracle — whose own capex cadence and AI training/inference roadmaps drive the order book.
Within the listed competitive set, AMD, Intel, and Broadcom each pursue overlapping but differentiated angles: AMD with discrete data-center GPUs in the MI series, Intel with foundry-plus-Gaudi accelerator combinations, and Broadcom with custom XPU programs sold into the same hyperscaler customers. Huawei HiSilicon's Ascend line sits in the China-domestic substitution segment that is structurally separated from NVIDIA's addressable market by export-control regimes. The relevant insight from the supply map is that NVIDIA's quarterly revenue trajectory is mechanically gated by upstream allocations — CoWoS capacity at TSMC and HBM stack output at the three memory IDMs — at least as much as by downstream demand.
What to watch
FY2027-Q1 earnings release (historically late May 2026) — will clarify whether Q4 DSI of 221 days normalises or extends
Korea Customs semi-export YoY print for May 2026 (early June release) — six-month acceleration trend continuation check
Quanta Computer May 2026 monthly revenue (mid-June 2026 release) — AI server ODM sell-through proxy
[2]Internal price summary — NVDA last 90 trading days— Share price moved from $186.23 on 2026-01-16 to $225.32 on 2026-05-15, +21.0% over 83 trading days
[3]Internal supply-chain map — NVDA upstream/downstream 1-hop— Fabless model dependent on TSMC, SK Hynix/Samsung/Micron, and ASE; concentrated hyperscaler downstream
Over the eight-quarter window, NVIDIA's quarterly revenue grew from $7,192 million in 2025-Q1 to a peak of $91,166 million in 2026-Q3 before settling at $60,922 million in 2026-Q4. That is roughly a 12.7× expansion at the peak versus the starting period. Operating income scaled even faster, from $2,140 million to $57,419 million at peak, reflecting fixed-cost dilution typical of fabless models when wafer and HBM allocations are pre-committed and incremental revenue carries near-software-like incremental margin. The operating margin moved from 29.8% in 2025-Q1 to 63.0% in 2026-Q3 before reverting to 54.1% in 2026-Q4. Net income tracked operating income closely, ending at $29,760 million in the latest quarter at $1.19 diluted EPS.
Two quarters interrupt the otherwise monotonic expansion. The first, 2025-Q4, shows revenue of $26,974 million but COGS of $11,618 million and operating income of just $4,224 million — implying a sharp margin compression that quarter consistent with an inventory-charge or write-down event embedded in COGS. The second, 2026-Q4, shows revenue down 33.2% sequentially against the Q3 peak, with COGS at $16,621 million and operating income normalising to $32,972 million. The DSI metric reinforces the second event: inventory days widened from 125 to 221 quarter-on-quarter, the highest reading in the eight-quarter window. Implied inventory rose from roughly $30.3 billion at Q3 (125 days × Q3 COGS run-rate) to roughly $40.4 billion at Q4 (221 days × Q4 COGS run-rate), a build of about $10 billion in a single quarter.
A DSI of 221 against the Q4 COGS run-rate is consistent with a major product-platform ramp — advance staging of wafers, HBM modules, and CoWoS-built systems for a next-cycle product — and is also consistent, though less probably so given the surrounding macro signals discussed below, with slower sell-through into hyperscale channels. The two interpretations are not mutually exclusive.[4]
Cash & Capital Cycle
The cash-flow statement is not in the input set, so the discussion is limited to working-capital proxies. The DSI re-expansion from 125 to 221 days, against a quarter in which COGS fell from $22.0 billion to $16.6 billion, means inventory dollars rose disproportionately to cost of revenue. In a fabless model, this typically reflects (i) pre-positioned wafer starts at TSMC for a next-platform launch, (ii) HBM3E/HBM4 stack inventory being staged ahead of customer pickup, and/or (iii) finished GPU server-module inventory awaiting CoWoS-OS or InFO_R packaging slot. None of these conditions is, by itself, a negative cash-cycle signal — the working-capital tie-up is a cost paid for optionality on a near-term ramp.
But the magnitude — an approximately $10 billion sequential build — is large enough that it will mechanically pressure free-cash-flow conversion in 2026-Q4 versus 2026-Q3, and any further DSI expansion in the next-quarter print would warrant re-examination. Operating-income dollar levels are sufficient that even a meaningful working-capital absorption need not signal financial stress; the question is one of cadence, not solvency. Reading the FY2026 10-K inventory note is the natural way to disambiguate the composition of the build.
Recent Events
Five SEC filings within the trailing 90 days appear in the disclosure feed. Three are low-impact 8-Ks announcing officer/director changes (2026-05-08, 2026-04-27, 2026-03-06). One is the FY2026 10-K filing on 2026-02-25, and one is the FY2026-Q4 earnings release on the same date. No high- or medium-relevance external news items where NVDA is the affected ticker appear in our 90-day filter. The cadence of three officer/director 8-Ks in roughly sixty days is higher than steady-state, but each is tagged low-impact; without further texture we treat the cluster as routine governance turnover rather than a signal in itself. The FY2026 10-K is the authoritative document for risk-factor disclosure, segment splits, and the full inventory note that would clarify the Q4 DSI build.[5]
Industry Context
The 1-hop supply-chain and macro indicators point to continued AI-accelerator demand strength heading into the next reporting period.
Korea's semiconductor export index (semi_export_403Y003_3091AA) shows a clear acceleration: -3.76% YoY in 2025-11, +7.26% in 2025-12, +22.48% in 2026-01, +19.78% in 2026-02, +35.64% in 2026-03, and +55.98% in 2026-04. The April 2026 reading of +56.0% YoY is the strongest in the six-month window and is consistent with a step-up in HBM and high-end logic shipments out of Korea — which in turn is consistent with NVIDIA, as the largest single end-customer of HBM, drawing on Korean supply at scale.[6]
Quanta Computer's April 2026 monthly revenue (released 2026-05-15) printed at NT$339,921 million, +120.7% YoY but -6.3% MoM. Quanta is the largest ODM exposure to NVIDIA's reference-design AI server platforms, so its monthly run-rate is a useful — if noisy — proxy for shipped AI server demand. A doubling-and-then-some YoY at the ODM layer corroborates strong sell-through; the modest MoM decline is consistent with normal lumpiness in server-system shipment timing and does not, on its own, suggest a turn in the cycle.[7]
DRAM spot indicators on 2026-05-16 are at elevated levels for high-bandwidth-relevant SKUs (DDR5 RDIMM 32GB at $960). Change-vs-prior is not reported in this snapshot, so the directional read is anchored on the export-index trend rather than the spot tick. The combination of accelerating Korea semi exports, robust Quanta AI-server revenue, and firm DDR5 RDIMM pricing collectively describes an industry context in which NVIDIA's elevated DSI is more plausibly explained by ramp-staging than by demand-side softness, though the FY2026-Q4 10-K inventory note would be needed to confirm the composition.[8]
Peer Benchmark
The two in-universe peers carry latest financials only for 2026-Q1, so the comparison is one-quarter-behind NVIDIA's latest print. Intel posted -$301 million operating income, -$821 million net income, and -$0.19 EPS on a DSI of 133. Broadcom posted $6,260 million operating income, reported net income of zero (likely a presentation artifact in the input table), and $1.14 EPS on a DSI of 43.
NVIDIA's 2026-Q4 DSI of 221 is materially above both peers. Broadcom's much-lower DSI of 43 reflects an ASIC business with substantially shorter wafer-to-revenue cycles and less HBM stacking on hand. Intel's DSI of 133 reflects an integrated device manufacturer with internal fab inventory exposure but is not directly comparable to NVIDIA's outsourced-stack profile. On the profit line, NVIDIA's $33.0 billion operating-income quarter dwarfs Broadcom's $6.3 billion and stands in stark contrast to Intel's loss. The peer set illustrates how far the AI-accelerator margin profile has diverged from the rest of the listed semiconductor universe in the current cycle. The peer comparison should be read as a context check rather than a relative-value statement.[9][10]
Watch List
The next-quarter FY2027-Q1 earnings release — historically a late-May print for NVIDIA — is the next major data point and will resolve much of the inventory-build ambiguity. The Korea Customs semi-export prints (typically released on the first business day of each month) and Quanta's monthly revenue (typically mid-month) will provide intra-quarter triangulation. The FY2027-Q1 10-Q inventory footnote will clarify whether the Q4 DSI build was raw-wafer staging, HBM-stack pre-positioning, or finished-system inventory. Officer/director 8-Ks should continue to be tagged low-impact unless a CEO/CFO-level change appears. CoWoS allocation commentary from TSMC's monthly revenue release and quarterly investor briefing is the upstream-supply check.
If Korean semi exports continue to accelerate and Quanta's May revenue holds the +100% YoY band, the ramp-staging interpretation of the Q4 inventory build gains weight. If, conversely, the YoY decelerates and DSI extends further in 2027-Q1, a re-read of the demand-absorption thesis would be warranted.
Research note based on public disclosures, prices, trade data, and supply-chain relationships. Hedged commentary is informational and not a solicitation to buy or sell any security.
— Broadcom 2026-Q1: OI $6,260M, NI $0 (table artifact), EPS $1.14, DSI 43
