Simon

I build open-source market data tools with AI.

a-stock-data 7.7k★ · TradingAgents 2.6k★ · global-stock-data 1.2k★

Not investment advice. I may hold positions

[Simon @linsizhen]

I shipped an AI multi-agent dashboard this week.

Then I cut the AI down to about 10% of what's on the screen. On purpose.

It's a daily-review board for short-term A-share traders. Every evening it pulls the limit-up pool, the consecutive-limit ladder, block-trade prints, sector money flow and theme attribution, and collapses an hour of manual work into one screen.

Five analyst agents, one judge agent. Standard setup.

But every number you actually read is computed, not generated.

The reason is boring: those readings are arithmetic. Median move of yesterday's strong names, follow-through rate, ladder distribution, market breadth. Each has exactly one right answer. Hand that to a model and it will probably get it right, or it will get it plausibly wrong. Nobody questions a number that's well formatted and looks reasonable.

So the split is:

▸ Hard metrics: pure computation, never touches the model, there the second you open it
▸ Five analysts + judge: the only job the LLM has is turning eight data sources into something a human can read

Unplanned benefit. When your key expires or the model rate-limits, 90% of the product still works. Built the other way round, a model outage is a blank page.

341 tests. Most of them aren't guarding against crashes. Crashes announce themselves. They guard against the other kind: renders fine, numbers look reasonable, conclusion is wrong.

Runs fully local. No API key needed if you already have a Claude or Codex subscription, it just drives the CLI you're logged into.

It doesn't pick stocks. Reviewing what already happened is factual work, and that never needed a recommendation.

Apache-2.0:
https://t.co/FiG4xBKWgB

[Simon @linsizhen]

Meta: free cash flow $784M, down 91%. Capex took 97.6% of operating cash flow. Cash halved to $15.5B. It still guided 2026 capex up.

Microsoft: capex +110%, free cash flow still $19.6B, capex at 65% of operating cash flow.

Revenue isn't the problem — Meta grew 28%, faster than Microsoft's 18%. Its operating income fell 8%.

You can fund +100% capex from profit, from cash, or from lenders. Only one of those compounds. So the number that breaks first isn't capex, it's the growth rate of capex.

Not investment advice. Company filings, 2026-07-29.

[Simon @linsizhen]

Microsoft and Meta both reported after the close. Both are spending record money on AI. Microsoft rose 3% after hours, Meta fell 8%. The gap is not about how much they spend.

Microsoft: revenue $90.0B (+18%), operating income $40.6B (+18%), adjusted EPS $4.74 against $4.24 expected. Azure grew 43% and passed $100B annually. Backlog rose 84% to $678B. Capex doubled to $35.8B for the quarter, $115.9B for the year, and free cash flow was still $19.6B. Capex took 65% of operating cash flow.

Meta: revenue $60.80B (+28%), which is faster growth than Microsoft. Operating income fell 8% to $18.78B. Margin went from 43% to 31%. Costs rose 55%, depreciation rose 46%. Capex $31.08B against operating cash flow of $31.86B, leaving free cash flow of $784M, down 91%. Cash fell from $35.9B to $15.5B.

One ratio explains both reactions. Capex as a share of operating cash flow: Microsoft 65%, Meta 98%, Alphabet 115%. Alphabet's free cash flow was negative $5.86B last quarter, and in June it priced an $84.75B equity raise to fund AI infrastructure.

AI capex has moved from being paid out of profit to being paid out of cash cushions and shareholder equity. Depreciation is the lagging part, and it does not negotiate.

None of them is cutting. Meta raised the floor of its 2026 guidance to $130–145B. As long as financing stays available, the orders hold. Amazon reports tonight and AWS is the cleanest test, since it sells compute without another business attached.

Not investment advice. Figures from company filings, 2026-07-29.

[Simon @linsizhen]

Nvidia -5% Monday. China's optics and PCB names fell
10-17% in a single session, now 35-50% off their highs.
Every chart screams "dip." I'm passing. Here's the full
reasoning — including the data that argues against me.

▸ THE BULL DATA IS REAL

I spent two days pulling every price I could find:

· B200 listed rents: $5.94/hr median — roughly 2x a
rough breakeven estimate
· Rental forward curve: August priced HIGHER than July
· China grey market: a DGX B300 server doubled in six
months to ~2.75x US retail
· DRAM contracts +13-18% QoQ; refurbished H100s sell
at just 15% below new

Physical demand for 2026 is not dead. That build is
contracted, and it will happen.

▸ WHAT ACTUALLY BROKE

The WSJ reported Nvidia is in talks to backstop ~$250B
of financing tied to OpenAI's datacenter buildout.
Read that twice: the seller guaranteeing the customer's
ability to buy the seller's product.

Telecom veterans have seen this movie. Lucent and
Nortel financed their customers' purchases in 1999.
Revenue looked great — until it didn't. Nvidia's CDS
just had its biggest one-day move on record. The
market is repricing the guarantor.

▸ THE LOGIC THAT MADE ME WALK

Every price I can track — rents, grey market, memory —
is a LAGGING indicator. It measures today's scarcity,
not tomorrow's returns. What broke last week is the
financing layer, and credit is a LEADING indicator.

And one simple thought experiment: if any big buyer
cuts capex, compute gets cheaper for everyone else —
so the others don't need to add. The "everybody adds
forever" assumption fails in both directions.

I have no edge on 2027 credit quality. And when you
have no edge, you don't need a position. There's an
old Chinese saying: a wise man does not stand under a
wall that is being load-tested.

▸ WHERE MY ATTENTION GOES INSTEAD

Humanoid robotics. Not because it's cheap — it isn't —
but because its driver is a product cycle, not a
credit cycle:

· During Monday's compute crash, humanoid supply-chain
names fell ~2% while compute proxies fell 10-17%.
One day proves little, but the market drew a line.
· Tesla's Optimus V3 reveal window is Jul-Aug; the
Fremont line targets ~70k units/yr annualized
· Unitree cleared IPO registration in 104 days — about
to become the first PROFITABLE humanoid maker to
go public
· 2025 global shipments: 13,318 units, +479% YoY

The layer I study is the bottleneck components —
harmonic reducers, planetary roller screws, dexterous
hands — where irreplaceability lives.

▸ METHOD

Same discipline as always: predefined entry bands, no
chasing, no averaging into falling knives. Watching is
also a position.

Disclosure: I hold a position in a Chinese humanoid
supply-chain company (harmonic drives). No position in
any compute name. Not investment advice.

[Simon @linsizhen]

Everyone paywalls options Greeks. I built a free one.

Full Greeks + IV + same-day 0DTE flow, straight from CBOE's
official feed. yfinance doesn't have Greeks. OpenBB's free
tier doesn't either.

Also in there: daily short volume for all 12,112 US symbols
(FINRA), a market-wide screener across 5,309 companies
(SEC EDGAR), and same-day insider + institutional filings.

13 layers, 30+ endpoints, zero API keys. One Python file,
MIT licensed.

Built with Claude Code. Repo link below.

[Simon @linsizhen]

━━━━━━━━━━━━━━━━━━━━
CXMT LISTS TOMORROW
20260727 · WHAT TWO MARKETS PRICE
━━━━━━━━━━━━━━━━━━━━

China's largest DRAM maker — world #4 — lists on the
STAR Market Monday. IPO price 8.66 CNY, or 579B CNY
(~$85B). Only 6.73% of the shares float on day one.

▸ TWO MECHANISMS, ONE NUMBER

Hyperliquid pre-IPO perp (xyz:CXMT)
$6.09 → 2.77T CNY (~$408B)
Polymarket day-1 closing-mcap ladder
median ~2.76T CNY

A leveraged perp and a set of binary event contracts.
Different traders, different mechanics, no shared order
book. They land 0.4% apart. The convergence is the
signal, not either number by itself.

▸ THE LADDER

≥2.0T CNY 88.5%
≥2.5T 65.5%
≥3.0T 35.5%
≥3.5T 21.5%

Look at the tail. The 3.0–3.5T bucket holds only 14%,
but everything above 3.5T holds 21.5%. The book is
pricing a break of 3T as a break to somewhere much
higher.

Also: the perp has de-rated 18% in eleven days, from
3.37T on pricing day to 2.77T tonight.

▸ THE WEEKEND CUT BOTH WAYS

Friday — Micron -7.0%, SK Hynix -8.3%, Samsung -7.6%,
SOXX -4.4%.

Saturday — Korea and the US announced $950B of memory
MOUs. SK Hynix: $750B of long-term supply to US buyers,
including a $500B+ Nvidia partnership and a 2GW Vera
Rubin + HBM4 datacenter for 2027. Samsung: $200B with
Broadcom, covering sub-2nm foundry and packaging. SK's
chairman relayed that Broadcom's CEO told him memory
demand would likely surpass available supply.

The perp went 2.81T → 2.77T across both. Professional
money read an 8% sector drawdown and a $950B demand
signal as roughly offsetting.

▸ THE PART NOBODY IS PRICING

2.77T is ~26x 2026 earnings — H1 guidance doubled, then
stripped of ~25% minority-interest leakage. The IPO
price was ~5x. Retail is anchoring on Micron's ~$1T cap
instead.

But the prospectus says the H1 surge is not sustainable.
2025 depreciation ran 4.6x adjusted net income. And the
29.5B CNY use-of-proceeds contains zero HBM — I grepped
all 401 pages. HBM is the option. DRAM contract pricing
is the earnings.

▸ METHOD

Pulled tonight: Hyperliquid /info, Polymarket gamma API,
the listing filing on CNINFO (Jul 24), Yahoo closes.
FX 6.79. Share count 66.881B. Event-contract depth is
thin (~$160k total) — read the ladder as a sentiment
scale, not a price.

No position. I applied for the IPO and was not
allocated. Not investment advice.

[Simon @linsizhen]

━━━━━━━━━━━━━━━━━━━━
📡 HBM HIRING RADAR
20260726 · BASELINE
━━━━━━━━━━━━━━━━━━━━

SK HYNIX vs MICRON

▸ SNAPSHOT

SK hynix America 42 openings
Micron (HBM roles) 193 openings

First edition = baseline. From here, every issue ships the delta.

▸ THE TREND

The leader is moving toward the customer.
The challenger is moving toward margin.

▸ THE DETAIL

SK hynix — 76% of US roles sit in San Jose. And they aren't memory jobs:
• 3D Stacked DRAM-on-Logic Design Engineer
• AI/HPC System Architect (x2)
• AI Memory Solution Architect
• Director, System-in-Package Architecture

DRAM-on-logic is base-die work. You don't hire four system architects in Silicon Valley to sell commodity die. That's a memory vendor refusing to be commoditized.

Micron — 193 HBM roles across six countries:
• Singapore (Fab 10A) — 54
• Richardson, TX — 47
• Taichung — 22 · Hyderabad — 17
• Jalisco, MX — 16 · Boise HQ — 8

Design in Texas, packaging in Singapore, 28 layout seats largely in Mexico. That's a cost curve being engineered.

▸ METHOD
Public ATS endpoints (Greenhouse + Workday), keyword-matched. Baseline edition — no month-over-month claim made.

[Simon @linsizhen]

━━━━━━━━━━━━━━━━━━━━
📡 HUMANOID HIRING RADAR
20260726 · 30-DAY DIFF
━━━━━━━━━━━━━━━━━━━━

TESLA OPTIMUS vs FIGURE

▸ SNAPSHOT (Jun 25 → Jul 26)

Tesla Optimus 231 → 282 (+22%)
Figure 101 → 122 (+21%)

Tesla opened 80 new reqs, net +51 — meaning ~29 closed or filled. This board is moving, not parked.
Figure posted 25 new reqs in the window.

▸ THE TREND

Same growth rate. Opposite bets.

Tesla is buying capacity.
Figure is buying deployment.

▸ THE DETAIL

Tesla's 80 new reqs:
• Manufacturing — 34 (42%)
• Tesla AI — 31
• Austin, TX — 24 of them
Production-hell playbook: build the line first.

Figure's 25 new reqs:
• Data Collection — 9
• BotQ Manufacturing — 3
• Commercial Operations — 3

But the titles are the real signal:
"Humanoid Robot Operator — Commercial Site Team"
"Apprentice Robot Service Technician"
"Fleet Coordinator" · "TeleOps Quality Support"

You don't hire site operators for robots that aren't on sites yet.

▸ ONE MORE THING

Figure opened AI Data Operations Manager roles in the UAE and Mexico City this month, plus its first Asia supply-chain seats.

Data ops is going offshore before the robots ship at volume.

▸ METHOD
Public ATS endpoints, identical query both months. Tesla exposes no post dates — new reqs inferred from ascending job IDs.

[Simon @linsizhen]

Most "free" US market data APIs stop at quotes.

No options Greeks. No short volume. No SEC filings.

So I built one that has them — 13 layers, 30+ endpoints, zero auth, no API key.

Every source labeled with its license tier.

https://t.co/1wB377UHQL

[Simon @linsizhen]

I built a-stock-data — an open-source A-share data toolkit for AI agents (Claude Code, Codex).

Launched May 2026, now 6.7k⭐ on GitHub.

Zero auth, 100% free: quotes, K-lines, research reports, fund flows, filings, limit-up boards, ETF options & more.

https://t.co/oKF2wpCYJK

[Simon @linsizhen]

I open-sourced Vibe-Research — the personal AI stock-research dashboard I use daily.

Deep China A-share data, plus US / HK / Korea. Quotes, valuation, financials, fund flows, news — all wired into your own AI. It never recommends a stock.

https://t.co/mal5q6dtWa

[Simon @linsizhen]

@pixelverse_xyz @Delphi_Digital @MeritCircle_IO @MechanismCap @BitscaleLabs @borgetsebastien @dingalingts Good!

[Pixelverse @pixelverse_xyz]

Pixelverse, the emerging game ecosystem and entertainment studio, raised $5.5M from top VCs and Web3 founders.

Investors include @Delphi_Digital , @MeritCircle_IO , @MechanismCap , @BitscaleLabs and more, along with angel investors like @borgetsebastien , @dingalingts , @graildoteth and @LucaNetz