What the last 30 days of S&P 500 8-K filings actually say about AI infrastructure spend — read straight from the filings, not the headlines.
The setup
Between April 14 and May 14, 2026, S&P 500 companies filed 572 8-Ks under items 8.01 (Other Material Events) or 2.02 (Earnings). I parsed every one of them — primary documents and all 902 exhibits — looking for one thing: where is real AI infrastructure spend being disclosed, and who’s named in the disclosure?
Out of 572 filings, 48 carried AI / LLM / NVIDIA signal. Of those, 15 were high-confidence — meaning a named counterparty plus a concrete figure or specific architecture commitment, not generic “we’re investing in AI” boilerplate. NVIDIA was named in 13 of the 15.
This is what those 15 filings actually said.
The headline read: capacity, in numbers
Five filings are worth quoting before anything else, because they’re the ones with hard figures attached — and the figures are not small.
- Amazon (AMZN): “one million+ NVIDIA GPUs to be deployed starting in 2026 … landed 2.1 million+ AI chips over the past 12 months.” AWS named OpenAI, Anthropic, Meta, and NVIDIA as customers in the same filing.
- AMD: 6 gigawatts of Instinct GPUs to Meta. A capacity number disclosed in gigawatts, not units.
- Microsoft (MSFT): $31.8B in quarterly capex — an ~$82.9B annualized run-rate, framed inside the “agentic computing era.”
- Alphabet (GOOGL): Google Cloud at $20.0B quarterly revenue, up 63% year over year, “led by enterprise AI infrastructure and AI solutions.”
- Cisco (CSCO): raised FY26 AI-infrastructure orders guide from $5B to $9B. A near-doubling, mid-fiscal-year.
If you’ve been reading “AI capex” as a vibes-based macro narrative, these are the numbers it actually reduces to.
The full top 15
| # | Ticker | Headline |
|---|---|---|
| 1 | NOW | ServiceNow + NVIDIA Enterprise AI Factory integration; Google Cloud Partner of the Year |
| 2 | SLB | NVIDIA “AI Factory for Energy” reference architecture (oilfield services) |
| 3 | CTSH | Cognizant AI Factory on Dell + NVIDIA; Gemini Enterprise Practice |
| 4 | BK | BNY on-prem NVIDIA SuperPODs + multi-year OpenAI deal |
| 5 | NXPI | NVIDIA robotics collab + eIQ Agentic AI Framework |
| 6 | AMZN | 1M+ NVIDIA GPUs in 2026; AWS deals with OpenAI, Anthropic, Meta, NVIDIA |
| 7 | PLTR | Rule of 40 = 145%; framing as AI-infra peer of NVIDIA, Micron, SK hynix |
| 8 | AMD | 6 GW of Instinct GPUs to Meta |
| 9 | COHR | NVIDIA $2B investment; optical networking for AI data centers |
| 10 | INTC | Xeon 6 selected as host CPU for NVIDIA DGX Rubin NVL8 |
| 11 | AMAT | Synopsys + NVIDIA materials-modeling collab |
| 12 | DLR | 9 GW pipeline / 6 GW data-center capacity, Meta-anchored |
| 13 | CDNS | Agentic-AI EDA platform leadership |
| 14 | GOOGL | Google Cloud $20.0B (+63% YoY), AI-infra led |
| 15 | MSFT | $31.8B quarterly capex; “agentic computing era” |
Behind the top 15 sit another 19 medium-confidence signals (TMO, CHRW, IRM, MRK, WDC, NWSA, CFG, NFLX, EFX, ROP, BKNG, GLW, GRMN, KLAC, BR, NI, GEN, MSI, CSCO) and 14 low-confidence — generic AI mentions without a named partner or figure. Useful as a watchlist, but not material on their own.
The thing nobody’s saying out loud
NVIDIA was named in 13 of the top 15 filings. Not as a vendor in passing — as the specific architecture, the specific partner, the specific reference design that the filer is anchoring on. Across hyperscalers (AMZN, GOOGL, MSFT), chip peers (AMD, INTC), data-center operators (DLR, COHR), enterprise software (NOW, PLTR, CDNS), services firms (CTSH), and into sectors that have no business showing up on an AI list (SLB in oilfield services, BK in custody banking) — the gravitational center is the same name.
That’s the read. Either NVIDIA is in fact the single counterparty most of the S&P 500 has chosen to build on top of, or it’s the name everyone has decided they need to invoke in their 8-Ks to be taken seriously. Probably both. The filings don’t distinguish.
The cross-vertical surprises
Three filings are worth a second look because the company doesn’t fit the “AI infrastructure” frame:
- Schlumberger (SLB) — oilfield services. Anchored its 8-K on the NVIDIA “AI Factory for Energy” reference architecture. Energy services using AI for subsurface modeling is not new; making it a named reference architecture in an SEC filing is.
- BNY (BK) — custody bank. Building on-prem NVIDIA SuperPODs and disclosed a multi-year OpenAI deal. Banks are not supposed to be infrastructure builders; this one is.
- Digital Realty (DLR) — the only filer in the window disclosing capacity in gigawatts (9 GW pipeline, 6 GW data-center capacity, Meta-anchored). When the unit of measurement changes from rack-count to gigawatts, the conversation has changed.
Caveats, called honestly
- 8-K text only. Earnings-call transcripts are not in scope. That’s typically where the richest forward-looking AI-capex commentary lives. A second pass against transcripts is on the list.
- Tesla (TSLA) had an H100-capacity-ramp chart in its filing that didn’t surface in the top 15 — no named counterparty in the adjacent text, so the ranker under-weighted it. Worth a manual read.
- One false positive in the scan: Garmin’s “Instinct” (the watch product, not the AMD GPU). Caught and removed.
- “Named in a filing” ≠ “deployed and producing revenue.” Some of these are aspirational architecture announcements. The figures (AMZN’s million GPUs, AMD’s 6 GW, MSFT’s $31.8B) are harder to walk back than the rhetoric.
Method
- Universe: S&P 500 constituents (503 names) sourced from Wikipedia, cross-referenced to the SEC’s
company_tickers.jsonfor CIK resolution. 503 of 503 mapped. - Window: 2026-04-14 → 2026-05-14.
- Forms: 8-K with items 8.01 (Other Material Events) or 2.02 (Earnings).
- Parsing: primary document plus every exhibit (key insight: 8-K cover sheets are mostly procedural — the substance lives in the EX-99.x exhibits).
- Signal detection: regex bundles for AI infrastructure spend, LLM integration, and NVIDIA / compute partnerships, with word boundaries and XBRL-token pre-stripping to kill false positives.
- Ranking: distinct bundles hit × specificity (named counterparty, dollar/capacity figure). Confidence labels assigned by hand on top.
- Volume: ~2,549 SEC requests at ~5.5 req/sec sustained, zero rate-limit responses, identified User-Agent.
Working data lives offline. If you want a specific filing read in depth, ask.
— AdelaideIndex Research