Skip to main content
AI Watch (nav: Markets → AI Watch) is a dedicated dashboard for the defining trade of this market. The premise: the AI complex can’t be judged from equity prices alone — the earliest signals live in credit (what default insurance costs on the borrowers funding the buildout), hardware markets (what compute and memory actually cost), and usage (whether inference demand keeps growing). AI Watch puts all of it beside the stocks. Several of these series exist nowhere else as charts. Their sources publish today’s value only — so auramarkets records them daily, and the charts deepen permanently from here.

AI Universe

The heart of the page: one module, two levels of depth, toggled in the header.
  • The Complex — five groups: Hyperscalers (now including SpaceX, Alibaba, and Tencent), Mag 7 adds, Semiconductors, Memory (SanDisk through Samsung, plus Seagate and Everpure), and Neo-clouds (CoreWeave, Nebius). Group membership follows Morgan Stanley’s AI-infrastructure taxonomy, with a few deliberate keeps (SanDisk; the Mag 7 adds group). SpaceX (SPCX) joined at its June 2026 IPO — Morgan Stanley classifies it as a hyperscaler.
  • Value Chain — the complete Morgan Stanley value chain: 8 layers (Owners/Operators → Semi Production → Processor → Server Components → Server → Network → Internal Power/Cooling → Power Supply), 33 segments, ~220 names including the full international supply chain — TSMC’s Taiwanese suppliers, the Japanese equipment makers, Korean memory, Chinese optics, and the global power/grid complex. Where a company appears in several segments — NVIDIA is in five — the repetition is the read: it marks the chain’s chokepoints.
Each level offers four layouts and shares the thematic map’s control language:
  • Table — The Complex shows every name with price, returns, forward P/E against live analyst consensus, the 30-day revision in that consensus, RPO (contracted backlog in Bwithitsquarteroverquarterchange,straightfromSECfilings),andtheMarketfocuscolumn(themetricscoverageactuallycites,fromthe[researchcorpus](/research/overview)+X).ValueChainshowsthe33segmentsasaggregaterowsundertheirlayerheaders,withbreadth(B with its quarter-over-quarter change, straight from SEC filings), and the **Market focus** column (the metrics coverage actually cites, from the [research corpus](/research/overview) + X). Value Chain shows the 33 segments as aggregate rows under their layer headers, with **breadth (% above the 50-day)** and each segment's **best/worst name over the week** inline — expand any segment to see its constituents. In both tables the **category bands carry their own aggregates**: equal-weight returns per group or layer, the group's median forward P/E and summed RPO (the Hyperscalers band totals ~1.9T of contracted backlog), so the strips read as summary rows, not dividers.
  • Treemap — tiles sized by cap class, colored by the selected measure; hover for detail, click for the standard stock pop-out.
  • Bars — segments/groups ranked by the selected measure, with min–max constituent whiskers showing dispersion inside each; click a bar to drill into its members.
  • RRG — relative rotation vs SPY (the same math as the thematic and factor RRGs): which groups/segments are leading, weakening, lagging, or improving.
And three measures across the full lookback set (1D → 3Y plus a custom start date): raw % change, z-score (the move against that entity’s own volatility — a +2% day means nothing for CoreWeave and everything for Microsoft), and MA breadth (% of members above the 20/50/200-day average). Recent IPOs (SPCX) show raw % until they accrue enough history for z-scores and the RRG. Company logos render on the treemap tiles, table rows, and drilled bars — international listings included. The camera icon in the header exports any view: download a PNG or copy the image to your clipboard (branded, ready to paste into a chat), and in table views also Copy CSV — the visible rows, expansion state included, ready for a spreadsheet.

Baskets

Equal-weight rebased indices that follow the level toggle — the five groups at The Complex, the eight chain layers at Value Chain, always against SPY — with the same control language as the sector monitor: a lookback dropdown (1D → 2Y, YTD, custom dates), a % vs Zσ toggle, and three views — ranked bars around a zero line, the rebased lines, and a correlation matrix of the baskets’ daily returns over the selected window (which baskets are actually the same trade right now, and which diversify). The divergence between them is the story — memory and semis versus flat hyperscalers tells you where the cycle’s profits are actually landing.

Universe movers

Beside Baskets: the top and bottom five single names across whichever level is active — all ~30 core names at The Complex, all ~220 at Value Chain. Each row carries the full company name, its layer · segment address, a centered-zero bar, and the % / Zσ pair (the toggle picks which ranks; names without enough history drop out of the σ ranking). Any lookback from 1D to 3Y plus a custom date, and a line view that plots the current ten movers’ rebased paths over the window. This is where the chain’s tails surface — the Japanese substrate maker up 24% on the day, or the neo-cloud down 13% — names that would otherwise hide inside segment aggregates.

Credit — single-name CDS

Daily 5-year CDS settlement levels for fourteen AI-complex names, from ICE’s clearing house. Quoted as a price per 100 par — lower = more stress. Two tiles lead the panel:
  • The MSFT − ORCL gap — Microsoft is the market’s safest tech credit; Oracle is funding AI capex with debt. The gap between their CDS is the cleanest daily read on whether credit believes the buildout pays for itself.
  • CoreWeave’s level — the pure-play AI-infrastructure borrower, quoted on the high-yield convention.
Names on different coupon conventions (100bp vs 500bp) aren’t directly comparable — the table tags the exceptions. And leveraged loans (where private AI-datacenter lending lives) are invisible to every public feed; the panel says so rather than pretending otherwise.

Hyperscaler bond spreads

The cash-bond complement to the CDS panel: a daily G-spread basket — what hyperscaler bonds yield over Treasuries — built from free primary sources (per-bond yields from iShares LQD/IGLB daily holdings files, the Treasury curve from FRED). ~147 bonds across Oracle, Microsoft, Amazon, Meta, and Alphabet; the headline line is the 20y+ long basket (the long end is where the AI-capex debate prices), with per-issuer lines toggleable — Oracle trading at multiples of Microsoft’s spread is the story. A grey static reference shows Goldman’s hyperscaler bond basket (GSUCHS30), recovered from a published chart, for back-history to February 2026; our own series accrues daily from August 2026.

Capex engine

Straight from quarterly SEC filings — fact, not consensus: trailing-twelve-month capex per hyperscaler, the aggregate capex-to-operating-cash-flow ratio (the “how much of the cash machine goes straight back out” number), each name’s free cash flow, and RPO — contracted backlog, the metric the market-focus poll ranked #1. Oracle’s negative FCF and its RPO running at multiples of annual revenue sit side by side here, which is the whole debate in two numbers.

Compute & memory

  • Cost of compute — H100 rental $/GPU/hr from three markets daily: RunPod (spot and on-demand), Shadeform (the neocloud median, with a live-availability count), and Vast (the marketplace floor). The spot-vs-on-demand spread and availability are genuine tightness signals no vendor publishes as a series.
  • Memory spot — daily DRAM (DDR5/DDR4), NAND wafer, and GDDR prices, rebased for comparison, with the trend arrows from the source. HBM has no spot market anywhere — its story shows up through the memory equities and contract-price headlines instead.

Token demand

Daily tokens routed through OpenRouter (about 19 months of history), with the China share of attributed tokens as a sub-panel — a usage-weighted read on whether Chinese open-weight models are taking inference share. Source: OpenRouter (openrouter.ai/rankings).
OpenRouter is a router, not the whole market — it undercounts first-party traffic (ChatGPT, Gemini) and over-represents open-weight models. The trend is the signal, not the level.

Taiwan risk premium

TAIEX ÷ KOSPI, rebased — both are AI-supply-chain markets, so the ratio isolates Taiwan-specific geopolitical discount from the common semis beta. A falling ratio = the Taiwan discount widening versus Korea.

News & wrap

Wire headlines tagged with universe names over the last 48 hours, plus the AI wrap — a nightly synthesis of the day’s AI-complex news in a handful of bullets, each naming its tickers.

Honest limits

The hardware and CDS charts start thin — recording began August 2026, and depth accrues daily (token history backfills to early 2025; capex and consensus data carry their own history). Forward P/E uses current-fiscal-year consensus, not a blended next-twelve-months. And the Market focus labels are a snapshot of the narrative, refreshed by re-running the poll — not a live feed.