The S&P is holding above its 50-day. The Nasdaq isn’t. Two main events are going to shape the direction of the week, earnings on Wednesday post close and the FOMC meeting in Jackson Hole, with Kevin Warsh delivering his first keynote on Friday at 10am est. Everyone is expecting good earnings for NVDA. On the last 4 reactions , the stock sold off after the print. Momentum or lack their of should provide good insight on the next direction of the AI trade, which is currently stalling. Nothing out of the ordinary here as we are at the end of August with lighter volumes and wider spreads. Though selling off a drop today remains elevated and so does the 10YR / 30YR rates at 4.69% and 5.21%. With those two macro features in play, it puts pressure on growth stocks.

SPY sits at 765.06, above the 20-day at 763.63 and comfortably above the 50-day at 752.30. Session VWAP at 764.10 — price is pinned right on it. That’s not weakness. That’s a market waiting for news.

Look at the volume bars. They are continuing to lighten up. Commitment to capital is a bit scarce heading into the Fed meeting.

QQQ is a different picture. 708.21, with the 50-day sitting overhead at 718.26 and the 20-day at 710.49. Both are resistance now, not support. Session VWAP at 706.90 is the only thing holding the low end together.

Read more

Read more

MSTZ, the 2x inverse MSTR fund: $10.30, down 97% all time. SOXS, Direxion’s 3x semiconductor bear: $42.08, down 6% on the day, its max chart printing six-figure “prices” in 2023 that are pure split-adjusted ghosts, price is down 99.99% since inception.. SSPC, the 2x short SpaceX ETF: −31.24% in a single session.

These funds get there fees without a care in the world for the investor/retail trader. No matter how terrible/toxic the products are. 9/10 of these inverse etfs will go down 99% over their life cycle.

Retail looks at these and concludes they go to zero, short them. Half right. The missing half is where the money is.

Decay is not a trade. It’s a discount on your entry.

The structural bleed is real and persistent — SOXS has run roughly ten reverse splits since 2011 — but it accrues slowly, and it will not save a badly timed short. MSTZ was down 86% since inception as of June 30. It was also up 271% over the trailing twelve months at that same momen…

Read more

Earnings Call Highlights

Amazon printed its first $200 billion quarter and the stock gapped 15.3% — its largest single-day earnings move in years against an implied move of roughly 6–7%. Now the part most coverage got wrong.

The $5.75 GAAP EPS number is not an earnings beat. It is a mark-to-model event. Q2 net income of $62.6 billion includes $53.4 billion of non-operating other income, and the 10-Q is explicit about where it came from: an upward adjustment of approximately $50.5 billion to Amazon’s nonvoting preferred stock in Anthropic, recorded as a Level 3 fair value measurement reflecting observable price changes from Anthropic’s latest funding rounds. The associated $18.2 billion tax provision is $17.7 billion deferred — non-cash.

Strip it out and the operating quarter still holds up on its own:

• Adjusted EPS $1.97 vs $1.82 consensus — an 8% beat, not a 214% one.

• Revenue $200.6B, +20% YoY, versus roughly $196.5B expected.

• Operating income $27.5B, +43%, margin 13.7% vs 11.4% a year ago.

Two items flattered operating income by about $1.2 billion: ~$640 million of IEEPA tariff refunds booked as a reduction to cost of sales in North America, and ~$600 million from fair-value remeasurement of energy derivative contracts, landing mostly in AWS. Both are one-time. Adjust for them and operating income is ~$26.3B — still a 37% increase.

AWS is a huge part of the story. Revenue of $42.2B, up 36.7%, the fifth consecutive quarter of acceleration and the fastest growth in 18 quarters. Operating income $16.6B on a 39.4% margin, up 645bps YoY (520bps ex-derivative gain). Backlog of $496B with a weighted-average remaining life of 6.4 years. AI run rate above $25B; custom silicon run rate above $25B, both growing triple digits.

Advertising did $19.8B, +26%. Third-party seller services $46.8B, +16%. North America margin 7.9%, International 4.1%.

Three themes that matter:

1 — Jassy pre-empted the capex question with a return-on-capital framework, not a defense. He split the spend into two capital cycles: data centers, which absorb capital two years before monetization but then produce 30+ years of use across five to six generations of servers without repeating the upfront cost; and servers and networking, purchased months before deployment, breaking even in a little under three years against five-to-six-year useful lives, with most AI capacity contracted for at least five years. His framing of the near-term pain: heavy CapEx and free cash flow headwinds until the data centers come online, and then revenue growth outpacing incremental CapEx “at some point.” He also said the quiet part directly — “If the demand isn’t there, we won’t spend the capital.” That is the option value the market paid for on Friday.

2— Demand is being described as supply-constrained through 2028. Capex goes to approximately $220 billion this year, up from $200 billion, with the increase attributed to memory pricing. Even at that level, Jassy said Amazon will not have enough capacity to meet 2026 demand, expects the same in 2027, and called the demand already booked for 2028 "striking." He raised the ceiling on AWS from a few hundred billion to "very possibly be a trillion-dollar annual revenue business."

• In the last couple of weeks the market has been punishing hyperscaler Capex spending, 1st with earnings, which preceded the liquidation of Situational Awareness, leading to a broad sell off and possible targeted attack on AI-Infrastructure related momentum names. Where separated itself was on the call, where flexibility was highlighted and the ability to abandon orders was voiced. If the demand for more servers and networks slows, they can claw back some of that capex. I believe markets have been pricing some of these names with the thesis that all of these Capex costs are sunk, which is not the reality, and Jassy framed that perfectly.

3 — The AI adoption curve is "barbelled," and the middle hasn't moved yet. One end is AI labs consuming enormous compute plus a handful of runaway consumer apps. The other is enterprises extracting cost avoidance from customer service and process automation. The middle — existing enterprise production workloads, most of which don't yet use inference pervasively — is, in Jassy's view, the largest absolute segment and still ahead. On margins, Olsavsky was careful: AWS profitability "isn't random," it's efficiency gains, capacity optimization and fixed-cost management, and it will fluctuate.

Trainium and the chips business

• Chips revenue run rate now above $25B, growing triple-digit percentages y/y. Note that AI revenue run rate is also cited as “over $25 billion” with triple-digit growth — the disclosure doesn’t clarify the overlap between these two figures. Don’t add them.

• Anthropic and OpenAI both have multi-year, multi-gigawatt Trainium commitments. Jassy called them the two leading AI labs in the world. The OpenAI commitment is the most strategically significant datapoint on the call — it’s validation from a customer with every incentive to use alternatives.

• Startup and enterprise adoption broadening: Neurorobotics and Odyssey named as new unicorn adopters, joining Twelve Labs, Descartes Labs, Poolside AI, Karakuri, Metagenomi, NetoAI, Splash Music, plus Uber and Pinterest.

• Third-party Trainium sales are being explored, not committed. On Nowak’s question, Jassy said an increasing number of customers want Trainium chips separate from AWS cloud, that they’re actively having those conversations, and that there’s “a real chance we’ll do that in the future.” No timeline, no ROIC framework given for it. Treat as free option, not a modeled line.

• Nvidia relationship deliberately preserved — they’ll keep making AWS the best place to run Nvidia, with customers expected to run Nvidia indefinitely. The choice/competition narrative is positioned as driving inference costs down.

• Graviton is arguably the quieter but stronger story: used by 98% of the top 1,000 EC2 customers, revenue commitments up nearly 3x q/q, and Graviton5 ramping nearly 2x faster than Graviton4 did. Given the AI→core CPU argument, this is the chip that monetizes the flywheel.

Sector Inflection

We are at the point in the AI infrastructure cycle where capex is no longer the *only variable that determines the stock reaction. Cloud revenue acceleration is. I also think the market is underpricing the flexibility of many of these contractual agreements for server/network infrastructure.

Competitor Landscape

Microsoft (MSFT) — $464.72, +3.0% on 7/31 Q4 FY26 revenue of $90.0B (+18%) beat the $87.6B consensus; non-GAAP EPS $4.74 vs $4.24. Azure grew 43% against 39–40% guidance and crossed $100B in annual revenue for the first time. Commercial RPO stood at $627B as of the prior quarter. Capex hit a record ~$41B and free cash flow fell 23%. The stock rose 8.9% after hours and added roughly $260B of market value in an evening.

Two caveats worth carrying: roughly $0.27 of the EPS came from discrete items including a $3.2 billion Anthropic gain — the same accounting phenomenon inflating Amazon’s and Alphabet’s prints — and the reported CY2026 capex figure was influenced by a classification change taking effect in FY2027. MSFT trades at ~25x trailing / ~23x forward and remains down roughly 12% over 52 weeks despite the pop.

Read more

Read more

Read more

Taiwan Semiconductor (TSM) headlines the semiconductor earnings season on Thursday, July 16, with the conference call at 2:00 AM Eastern. Shares trade around $421/58 — up strongly on the year, roughly 10% below the 52-week high, and about 2% lower into the print after a bearish options-flow session Monday. Consensus rating is Strong Buy. Market cap sits near $2.27 trillion.

Here is the setup nuance that reframes the entire trade:

Revenue is already public. TSMC reports monthly sales, and Monday’s June release locked in the quarter at NT$1.27 trillion (~$39.6B), +36% YoY — at the top of guidance and slightly ahead of consensus. Thursday is a margin & guidance event, not a revenue event.

The options market is pricing a full-quarter surprise, but roughly 60% of the information — the top line — is already out. What is left to move the tape is gross margin, EPS flow-through, and above all the Q3 guide and the full-year outlook.

Implied volatility & the expected move

The options market is pricing an implied move of roughly ±5–6% into Thursday — Benzinga Pro shows ~6%, while Cboe data imply a 50% chance of a move greater than 5.2%. Put/call is running below historical norms (slightly bullish positioning), though bearish flow appeared Monday as traders hedged. Two things matter here: TSM historically realizes less than its implied move on earnings, and it reliably sees post-print IV crush. With revenue already known, the unknown left to price is smaller than the straddle implies — a setup that favors premium sellers unless the guide surprises hard.

The estimates

Consensus lands near $3.80 per ADR (range $3.61–$4.19), up 53–55% year over year from $2.47, against a full-year 2026 view of ~$15.91 (+49%). On revenue, the already-reported ~$39.6B is +36% YoY and +10% sequentially, topping the high end of the company’s own $39.0–40.2B guide. The bar: TSMC has beaten in each of the last four quarters, average surprise +8.3%, and has reportedly pushed advanced-node prices 5–10% higher — both of which skew EPS toward another beat.

Guidance: the actual event

This is where the move is made. Watch four things:

• Q3 revenue & gross-margin guide — Q2 GM was guided 65.5–67.5%; the Street models ~45% EPS growth for Q3.

• Full-year raise — Citi expects TSMC to lift both its 2026 revenue-growth outlook (currently “above 30%”) and long-term targets, and placed a 30-day upside catalyst watch on the stock. A move toward mid-30s% growth is the bullish trigger.

• Capex — management has guided toward the high end of $52–56B; any raise above $56B is an incremental bullish read-through for the whole AI-capex chain.

• N2 margin bridge — the 2nm ramp plus overseas fabs carry 2–3% gross-margin dilution, widening to 3–4%. The framing of that path is the swing factor on the call.

Track record favors the bulls on process: TSMC guides conservatively and beats. Q2 just topped its own range.

Read more

The following chart is a great/reliable macro indicator I use called Chillax MA on TradingView , showing a potential choppy market head on the daily, when we are flashing light green in signaling momentum is on and progressive exposure is warranted, notice the transition from March to April, where we transitioned from black to Green, marking the the beginning of the largest semiconductor run ever. I am not saying the indicator is perfect, I am saying it is a decent way to know when more risk on is warranted (transition from bear/chop—→ bull cycle) and when you want to be more cash or even looking for shorts.

Read more

Decent recoveries along after a sharp multiday sell off where back to back selling days resulted in many breaks of momentum leaders like and the closing below the 50 day simple moving average (SMH) for the first time since early April. Though we have a nice bounce on our hands I don’t think we are completely out of the woods yet and there is resistance in the semi space ahead. Another positive sign this week is from reclaiming its 200sma with authority. When largest AI -earner in the world reclaims some strength on a China-supply deal it instills some confidence. SOXL bouncing another 15% today is also what you could expect after a 50% drawdown from highs. It’s resting right against the 50day sma here at 200.79 and the coming days will provide more clarity. Ultimately I think there will be a lot of chop going forward inthe semi-space and it’s not where i’m most interested. These stocks have made multiple hundred/thousand percent moves over the past few months/year and some sideways price action would be healthy for digestion.

SK Hynix $SKHY has filed its Form 8-A ahead of its U.S. listing tomorrow
Bloomberg is reporting shares are expected to be priced at $149 per ADR, or $1,490 per share. The way this trades on Friday and over the next week will provide a lot of info given investor expectations of the memory super cycle, we will see.

Where is Money Rotating/Next Plays.

Read more

Read more

CEREBRAS SYSTEMS (CBRS) — INSTITUTIONAL EQUITY RESEARCH BRIEF

Data vintage: Compiled May 14, 2026 (IPO day). Financial data sourced from amended S-1/A filed April 2026 (FY2025 figures) and IPO pricing announcement May 13, 2026.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. COMPANY SNAPSHOT

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

• Full name: Cerebras Systems Inc.

• Ticker / Exchange: CBRS / Nasdaq Global Select Market

• Sector / Industry: Information Technology / Semiconductors & AI Infrastructure

• IPO price: $185.00 (priced May 13, 2026; 30M Class A shares; $5.55B gross proceeds; ~17% above the revised $150–$160 marketed range)

• Fully diluted valuation at IPO: ~$56.4 billion

• Pre-IPO secondary market reference (Hiive): ~$187.53 (May 12, 2026), implying day-one pricing is roughly in line with secondary market

• Founded: 2016, Sunnyvale, CA. CEO/co-founder Andrew Feldman.

• Float: 30M shares (+4.5M underwriter greenshoe option). Float is <10% of fully diluted share count — extremely thin, a critical setup feature.

Business description: Cerebras designs and manufactures wafer-scale AI processors (flagship: Wafer-Scale Engine 3 / WSE-3 — a single chip the size of a dinner plate with 4+ trillion transistors, claimed to be 58× larger and up to 15× faster on inference than leading GPU solutions). The company sells (i) integrated AI supercomputer hardware systems (CS-3) to enterprises, governments, and AI labs, and (ii) cloud-based inference services running on its own infrastructure. The strategic pivot from pure hardware sales toward operating an inference cloud is central to the bull story — and to the risk profile.

FY2025 revenue mix:

SegmentRevenue% of totalHardware (CS-3 systems)$358.4M70.3%Cloud & other services$151.6M29.7%Total$510.0M100%

Customer concentration (per S-1/A):

• Mohamed bin Zayed University of AI (UAE) — 62% of 2025 revenue

• G42 (UAE) — 24% of 2025 revenue (down from 85% in 2024)

• Combined ~86% to two UAE-affiliated entities in 2025

• OpenAI deal ($20B+ over 3 years) and AWS partnership begin contributing materially in 2026+

Geographic exposure: Heavily skewed to UAE in 2025; expected to rebalance toward US (OpenAI, AWS) in 2026–2028.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

2. FUNDAMENTAL ANALYSIS

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

INCOME STATEMENT (per April 2026 S-1/A)

MetricFY2022FY2023FY2024FY2025Revenue$24.6M$78.7M$290.3M$510.0MYoY growthn/a+220%+269%+75.7%Gross margin~11.7%~33.5%~41% (H1)mid-40s% (est.)GAAP net income/(loss)($177.7M)($127.2M)($481.6M)+$237.8MNon-GAAP net income/(loss)n/an/a($21.8M)($75.7M)

Critical accounting flag: The headline FY2025 GAAP net income of +$237.8M is misleading. It includes large non-cash gains tied to the change in fair value or extinguishment of forward contract liabilities (related to the OpenAI warrant/loan structure). Stripping these out plus $49.8M of stock-based comp, the company posted a non-GAAP net loss of ~$75.7M, ~247% worse YoY. There is also a GAAP operating loss of $145.9M in 2025. The operating business is still unprofitable. This is the single most important fundamental nuance in the entire filing.

• EPS history vs. consensus: Not applicable — no public consensus exists pre-IPO.

• Quarterly beat/miss history: Not applicable — first reportable quarter will be Q2 2026.

BALANCE SHEET

• IPO proceeds: ~$5.55B gross (~$5.2B net after underwriting fees) — pro forma cash will be substantial.

• Pre-IPO cash: Bolstered by $1B Series H (Feb 2026) and $1B working-capital loan from OpenAI (Dec 2025, at 6% interest, secured by warrants for 33.4M shares).

• OpenAI warrant overhang: Up to ~10% of fully diluted shares (~$5B value at IPO price) issuable to OpenAI as the relationship scales — material dilution risk if fully exercised.

• Remaining Performance Obligations (RPO): $24.6B as of S-1/A filing — provides multi-year backlog visibility (primarily OpenAI + AWS).

• Net debt/EBITDA: Not meaningful — company is operationally unprofitable on a non-GAAP basis.

CASH FLOW

Cerebras has not historically generated material operating cash flow. Expect heavy capex through 2026–2028 to build out OpenAI’s 750MW commitment (requires Cerebras to invest in data centers, not just ship boxes). The shift to a cloud/inference operator model transforms the capital intensity profile from a relatively capital-light fabless model to an infrastructure-heavy hyperscaler-adjacent business. FCF yield: materially negative for the foreseeable future. No dividend; no buyback (capital is being consumed, not returned).

GROWTH METRICS

• Revenue 3-yr CAGR (2022–2025): ~174% (off a very small base)

• 2026 implied trajectory: With OpenAI $20B over 3 years (~$6.7B annualized) and AWS ramp, consensus-style estimates point to $1.5–$3B+ in 2026 revenue, though this depends entirely on OpenAI’s 750MW deployment timeline.

• Sell-side coverage: None at IPO; expect initiations from Morgan Stanley, Citi, Barclays, UBS, Needham, TD Cowen 25 days post-IPO under the syndicate quiet period.

• Buy/Hold/Sell distribution: N/A at IPO.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

3. VALUATION & RELEVANT MULTIPLES

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

At the $185 IPO price with $56.4B fully diluted market cap:

MultipleCBRSNotesEV / TTM Revenue~110xOn $510M 2025 revenueEV / NTM Revenue (est.)~25–35xAssuming $1.5–$2.0B 2026 revenueEV / 2027E Revenue~10–15xHighly sensitive to OpenAI rampP/E (NTM)N/MNon-GAAP loss; GAAP figures distortedP/BN/MDistorted by warrant accountingFCF yieldNegativeHeavy capex phase

PEER COMPARISON

CompanyTickerMkt CapFY26E Rev GrowthFwd P/ENotesCerebrasCBRS$56B+200%+ (est.)N/MPure-play inference; ~110x TTM salesNVIDIANVDA~$4T+~50–65%~32–36xDominant; CUDA moat; cheapest of the four on fwd P/EAMDAMD~$400B+~32%~32–40xMI300/MI350 ramp; OpenAI deal already played outBroadcomAVGO~$1.5T+~52%~31–41xCustom ASIC leader; software resilienceTSMCTSM~$1.4T~30%~24xFoundry — the picks-and-shovels play

Premium/discount analysis: CBRS prices at ~110x TTM revenue vs. NVDA at ~19x forward P/S and ~25–30x TTM P/S at the peak of its hypergrowth. Even adjusting for Cerebras’s higher growth rate and inflection narrative, the multiple embeds heroic 2027–2028 assumptions. The $24.6B RPO does provide visibility, but RPO is not revenue, and the OpenAI contract carries non-trivial cancellation/renegotiation optionality given the warrant structure. Justification for the premium rests on: (1) scarcity value as the only pure-play AI inference IPO of 2026, (2) circular-economy halo from OpenAI deal mirroring AMD’s 3x rally after its OpenAI deal, and (3) thin float driving technical scarcity. The valuation is not supported by traditional comparables; it’s a sentiment-and-scarcity trade.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

4. TECHNICAL ANALYSIS

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚠️ Material limitation: CBRS opens for trading today (May 14, 2026). No moving averages, no 52-week range, no RS lines, no ADV, no institutional 13F data exist yet. The framework below sets expectations for setup conditions to watch.

• Reference prices:

• IPO price: $185.00

• Hiive secondary (May 12, 2026): $187.53

• Polymarket day-one close consensus: $50–70B market cap (~$165–$230/share equivalent)

• 52-week range: Not yet established.

• Volume: Will likely see >$3–5B+ in first-day notional, given oversubscription (~20x) and float scarcity.

• Float characteristics: 30M Class A shares (~5–8% of fully diluted). Thin float + heavy oversubscription = high day-one volatility and potential for explosive moves in both directions.

• Short interest: Zero at open. Borrow will likely be very tight/expensive in the first 30–60 days; meaningful short interest typically builds 2–3 weeks post-IPO.

• Lock-up expiration: Standard 180-day lock-up expires ~mid-November 2026 — a material structural overhang. Insiders (Benchmark $3.3B, Fidelity $3.8B, Foundation Capital $2.8B, Eclipse $2.5B, Feldman $1.9B) hold tens of billions in paper value.

• Options market: No listed options at debut; CBOE typically lists 5+ trading days post-IPO. Watch for elevated IV at listing.

• Institutional ownership: Will populate in 13F filings beginning mid-August 2026 (for Q2). Initial institutional allocation skewed toward the four lead underwriters’ top accounts.

Read more

$GBTG — General Catalyst-backed Long Lake nearing a deal for Amex GBT at ~$9.50/share, a ~60% premium to Friday’s close. Announcement could land as soon as today. Read: Clean take-out arb. Spread to deal price is the trade — watch for confirmation headline.

$AMD — HSBC downgrades to Hold, PT $340. Cites multiple expansion from 19x to 33x 2027E P/E with capacity constraints capping near-term EPS upside. Read: Valuation call, not a thesis kill. Watch how the semis complex digests it into the open — semis have been our anchor sector.

$CRCL / COIN — Crypto tape ripping on long-awaited stablecoin yield text. CRCL +19.36% to $119.00, COIN +7.57%. The new language — first reported by Punchbowl News — bars platforms like Coinbase from paying customers yield on idle stablecoin balances, but preserves rewards tied to bona fide transactions, liquidity and market-making, and collateral posted against trades or loans. Activity-based yield lives; passive yield doesn’t.

This matters because Circle and Coinbase share a long-term commercial distribution deal on USDC and have leaned on interest payouts as a core growth driver for years. The compromise removes the cliff-risk scenario — yield gets banned outright — that had been pricing into both names. Committee markup possible week of May 11, same week as Circle’s Q1 earnings. Read: Catalyst-on-catalyst into earnings. This is the cleanest regulatory tailwind the stablecoin trade has seen in months — but understand what it actually says before chasing. The bill draws a hard line between “yield on idle balances” (gone) and “rewards for using the coin” (preserved). That’s a structural win for transactional volume, less so for treasury-style USDC parking.

Read more

Q1 Revenue: $26.9M (+39% YoY, beat $26.22M est.) | Non-GAAP EPS: -$0.01 (beat -$0.04 est. by 75%) | Gross Margin: 29.9% (vs. -6.1% a year ago) | InP Backlog: >$100M (all-time record) | Close: $96.00 (+21.2%)

Earnings Call Highlights

AXT printed its cleanest quarter in years. Non-GAAP EPS of -$0.01 crushed the Street’s -$0.04 consensus by 75%, and revenue of $26.9M (+39% YoY) edged the $26.22M estimate. More importantly, the trajectory reads right: gross margin swung from -6.1% to 29.9% in twelve months — a 3,600bp reversal driven by InP mix-shift and volume absorption. This is not a miss-the-bar-by-less story. This is a company hitting escape velocity.

Q2 guidance is the real headline. Management guided Q2 non-GAAP EPS to $0.06–$0.08 — the first profitable quarter in this cycle — against a Street that expected a $0.01 loss. The $34M revenue floor (permits already in hand or not required) is conservative anchoring, with management explicitly flagging “even significant upside” if additional permits clear. InP backlog crossed $100M for the first time. The only binding constraint on this story is export permit timing from Beijing A few interlocking reasons, all stemming from the same core problem: AXT’s entire manufacturing operation runs through a Chinese subsidiary it doesn’t fully control.

Read more

Most people are asking AI the wrong kind of question. Here’s what changes when you stop asking and start architecting.

My last piece introduced a megaprompt that turns Claude into an institutional-grade equity research analyst. The response surprised me — not just in volume, but in the specific question that kept appearing in the comments and my inbox:

Why does it actually work so much better?

It’s a fair question. You’re using the same model. The same interface. The same underlying intelligence. And yet the output feels categorically different — more rigorous, more balanced, more useful. The gap between a generic prompt and a well-architected one isn’t incremental. It’s qualitative.

So today I want to get into the mechanics. Not the prompt itself, but the principles underneath it. Because once you understand why structured prompting works, you stop thinking about individual prompts and start thinking about how to architect AI interactions for any domain you care about.

The Vending Machine vs. The Colleague

Here’s the mental model most people bring to AI: they treat it like a vending machine.

Insert a question. Press a button. Receive an output. If the output isn’t great, rephrase the question and try again.

This isn’t wrong, exactly. It works for simple, well-defined tasks. But it fundamentally misunderstands what these models are and how they generate responses.

A large language model isn’t retrieving a pre-formed answer from a database. It’s making a probabilistic prediction — at each step — about what the most contextually appropriate next token is, given everything it has processed so far. Which means the context you provide doesn’t just inform the answer. It shapes the entire distribution of possible answers the model is choosing from.

The vending machine metaphor fails because it implies a fixed output for a fixed input. The better metaphor is a colleague.

When a colleague walks into your office and you say “what do you think about this company?” — their answer depends enormously on who they are, what their job function is, what frameworks they’ve internalized, and what kind of output they think you’re looking for. A CFO will answer differently than a salesperson. A bear will answer differently than a bull. Someone who knows you want a two-page memo will answer differently than someone who thinks you want a thirty-second verbal summary.

Structure is how you tell the model who to be, how to think, and what to produce. And that changes everything.

The Four Layers of Structured Prompting

When I break down why the equity research megaprompt works, it comes down to four distinct layers — each doing a different job.

Layer 1: Persona

The prompt opens with this:

“You are a senior equity research analyst at a top-tier hedge fund with deep expertise across fundamental analysis, technical analysis, sector dynamics, and macro frameworks. You operate with the rigor and precision of a CFA charterholder and the market intuition of a seasoned portfolio manager.”

This isn’t window dressing. Persona assignment is the highest-leverage single intervention you can make in prompt design. Here’s why.

Language models are trained on vast corpora of human-generated text. That text was written by people with different roles, different expertise levels, and different purposes. When you assign a specific professional identity, you’re not just telling the model to “act like” someone — you’re activating a specific region of the model’s learned associations. The vocabulary changes. The frameworks surface. The epistemic standards shift.

Ask Claude “what do you think about Tesla?” and you get a generalist summary. Tell Claude it’s a CFA-caliber hedge fund analyst and ask the same question, and it starts reaching for EV/EBITDA multiples, FCF yield, short interest as a percentage of float, and management credibility signals. The model knows these things. Persona is what summons them.

The specificity matters enormously here. “You are a financial analyst” is far weaker than “you are a senior equity research analyst at a top-tier hedge fund.” The more precisely you define the role, the more precisely the model can calibrate its output register.

Layer 2: Operating Principles

Persona tells the model who it is. Operating principles tell it how it thinks.

The equity research prompt includes lines like:

• “You present both a bullish and bearish case for every stock with equal rigor, regardless of your own lean”

• “You flag data quality issues, accounting irregularities, or information gaps explicitly”

• “You never give direct buy/sell recommendations — you present the evidence and let the user form conviction”

• “You cite sources, data vintage, and confidence level where relevant”

Each of these is doing something specific.

The symmetry rule — equal rigor on bull and bear — is probably the single most important quality control mechanism in the entire prompt. Without it, AI outputs drift toward confirmation of the implied sentiment in the question. Ask about a hot stock and you get optimism. Ask about a struggling one and you get pessimism. The model is pattern-matching to how humans talk about these things. The symmetry rule breaks that pattern and forces intellectual honesty by design.

The data quality flagging instruction does something equally important: it converts a potential failure mode (hallucination) into a visible signal. A model that’s been told to flag data gaps and uncertain sources will surface its own uncertainty rather than papering over it with confident-sounding prose. That’s not just honest — it’s usable. You know where to verify.

The no-recommendation rule keeps the analyst in its proper epistemic lane. Investment decisions involve personal circumstances, risk tolerance, time horizon, and position sizing that the model can’t know. By explicitly prohibiting recommendations, the prompt forces the output to be analytical rather than advisory — which paradoxically makes it more valuable, because it gives you the material to form your own conviction rather than an assertion to accept or reject.

Layer 3: Output Specification

This is where most people underinvest, and where the equity research prompt does the most structural work.

The prompt specifies nine numbered sections, each with a detailed list of required data points, framing questions, and sub-components. This does four things simultaneously:

It creates completeness pressure. When the model knows it needs to populate a peer comparison table, it has to find comparable companies and retrieve their multiples. When it knows it needs a debt maturity schedule, it has to go find one. You can’t generate a plausible-sounding peer table with made-up data — the structure makes hallucination harder to sustain.

It forces sequencing. The nine sections aren’t arbitrary — they follow the natural information hierarchy of investment analysis. You understand the business before you analyze the financials. You understand the financials before you assess the valuation. You understand the valuation before you build the bull and bear case. Each section conditions the next. This is how good analysts actually think, and the structure encodes that reasoning sequence.

It produces a comparable output. Every time you run the prompt on a new ticker, you get the same nine sections in the same order. That comparability is underrated. It means you can read two research briefs side by side and compare like with like. The structure becomes an analytical interface.

It signals quality standards. A model that knows it needs to produce “NTM P/E vs. sector avg vs. 3-year historical avg” is going to search for, retrieve, and present data differently than one asked to “discuss the valuation.” The specificity of the requirement is a quality signal that cascades through the output.

Layer 4: Behavioral Guardrails

The final layer is the set of rules that govern what the model won’t do and how it handles edge cases.

“If data is unavailable or unverifiable, say so explicitly — never fabricate figures.”

“Always state the data date or source vintage.”

“Aim for depth and precision over brevity.”

These aren’t just stylistic preferences. They’re quality control mechanisms that address the specific failure modes of AI-generated research: hallucination, false precision, surface-level coverage, and misplaced confidence.

The “never fabricate” instruction works in conjunction with the data quality flagging from Layer 2. Together they create a model that self-reports its own uncertainty rather than filling gaps with plausible-sounding fiction. The source vintage requirement forces the model to contextualize its data rather than present it as timeless truth. The depth-over-brevity instruction counters the model’s natural drift toward summarization.

The Compounding Effect

Here’s what makes structured prompting non-linear in its impact: the four layers interact with each other.

A well-defined persona makes the operating principles more coherent — a hedge fund analyst naturally thinks in terms of asymmetric risk/reward, so the symmetry rule and the no-recommendation rule feel like extensions of the role rather than arbitrary constraints. The output specification gives the operating principles something to act on — the symmetry rule is enforced by having separate, explicitly labeled Bull Case and Bear Case sections that the model has to populate with equal rigor. The behavioral guardrails reinforce the persona — a CFA charterholder doesn’t fabricate data or omit source dates.

Each layer amplifies the others. The total effect is dramatically greater than the sum of the parts.

Read more

Most people fire off prompts like “what do you think about Tesla?” and get back something that reads like a Wikipedia summary with a sprinkle of optimism. Helpful, maybe. Institutional? Not even close.

But here’s what’s possible with the right architecture: you can configure Claude to think, write, and reason like a senior equity analyst at a top-tier hedge fund — covering every dimension a real investment committee would demand before committing capital.

That’s exactly what this megaprompt does. And today I want to break down why it works, what it produces, and who it’s genuinely useful for.

What Is a Megaprompt, and Why Does It Matter?

A megaprompt isn’t a question. It’s a persona + process + output specification bundled into a single system-level instruction.

Where a regular prompt says “analyze this stock,” a megaprompt says:

• Here’s who you are (a CFA-caliber analyst with hedge fund sensibilities)

• Here’s how you think (intellectually honest, data-driven, equally rigorous on the bull and bear case)

• Here’s exactly what you produce (9 structured sections covering fundamentals, valuation, technicals, catalysts, sector dynamics, and risk/reward)

• Here are the quality standards (cite your sources, flag data gaps, never fabricate figures)

The result isn’t just a better answer. It’s a fundamentally different class of output.

What the Prompt Actually Generates

When you drop a ticker into this framework, you get back something that looks less like a chatbot response and more like a Bloomberg equity research note. Here’s what’s covered:

1. Company Snapshot — The facts. Market cap, enterprise value, float, business model, revenue mix, geographic exposure. No fluff, just orientation.

2. Fundamental Analysis — Three years of income statement data, balance sheet health, cash flow quality, FCF yield, dividend and buyback activity, and analyst consensus estimates. All labeled by fiscal year or TTM/NTM so you know exactly what you’re looking at.

3. Valuation & Multiples — NTM P/E, EV/EBITDA, EV/Revenue, P/FCF, PEG, and Price/Book, all benchmarked against the sector average, the stock’s own 3-year history, and a peer comp table of 3–5 direct competitors. The prompt forces an explicit answer to the premium/discount question — and demands reasoning for it.

4. Technical Analysis — Price vs. the 20, 50, 100, and 200-day moving averages. 52-week range positioning. Volume analysis. Support and resistance levels with rationale. Short interest as a percentage of float (and days to cover). Institutional ownership changes from 13F filings. Options market signals where notable.

5. Recent Catalysts & News — A ranked summary of everything material in the last 30–90 days: earnings, guidance changes, analyst upgrades/downgrades, insider activity, product launches, M&A, regulatory developments. Upcoming known catalysts are flagged explicitly.

6. Sector Dynamics — A structured headwinds/tailwinds framework covering industry growth trends, competitive positioning, macro conditions, secular themes (AI, electrification, reshoring, GLP-1, etc.), regulatory environment, and execution risks.

7. Bull Case — A proper long pitch: one-sentence thesis, 3–5 evidence-based reasons for outperformance, what the market is missing, a base case and bull case 12-month price target with methodology and implied upside, and the milestones that confirm the thesis.

8. Bear Case — A proper short pitch with the same rigor: core short thesis, 3–5 reasons for underperformance, what the market is ignoring, a bear case price target with downside, and the scenarios that cause a de-rating.

9. Risk/Reward Summary — Asymmetry framing, a 1–5 quality rating across three dimensions (Fundamental Quality | Technical Setup | Catalyst Clarity), the one or two factors that determine which case plays out, binary risk events on the horizon, and position sizing considerations.

Why This Works: The Principle Behind It

The reason this megaprompt generates dramatically better output than a generic query comes down to one insight:

AI models don’t just respond to questions — they respond to context.

When you give Claude a specific professional identity, a disciplined analytical framework, explicit output standards, and quality control rules (like “never fabricate figures” and “always cite data vintage”), you’re doing something powerful: you’re constraining the model’s output distribution toward the space of high-quality professional work.

You’re also forcing symmetry. Notice that the prompt explicitly requires an equally rigorous bull and bear case, regardless of the analyst’s lean. This is how real investment committees work — you don’t get to stand up and pitch a long without someone in the room stress-testing the short. The megaprompt bakes that intellectual honesty in by design.

The other key mechanism is structure as quality signal. When the model knows it needs to populate a peer comp table, it has to find comparable companies and retrieve their multiples. When it knows it needs to state data vintage, it has to track sourcing. When it knows the bear case needs to be as compelling as the bull case, it can’t phone it in. Structure creates accountability — even for AI.

Who Is This Actually Useful For?

Let me be direct about who gets the most value here:

Retail investors who want institutional-grade research — Most individual investors make decisions based on headlines, Reddit threads, or a quick glance at a P/E ratio. This framework surfaces the same dimensions that professional analysts look at: FCF quality, debt maturity schedules, short interest as a coil vs. a warning sign, what the market is pricing in vs. what could surprise. That’s genuinely edge-enhancing.

Finance students and early-career analysts — This is essentially a curriculum for how to structure equity research. The nine-section format mirrors how real sell-side and buy-side research is organized. Using it repeatedly teaches you what questions to ask and why they matter.

Portfolio managers doing rapid screening — A PM who needs to get up to speed on a new name quickly doesn’t want to read a 40-page initiation report. They want the bull/bear framing, the key multiples vs. peers, the technical setup, and the upcoming catalysts — in that order. This delivers exactly that.

Read more

Read more

Read more

Read more

This is earnings season. And if you’re not treating it as the single most important setup generation window in your trading calendar, you’re leaving asymmetric opportunity on the table.

Let’s break down why earnings season matters, how a single print and eps guide can alter the trajectory of a stock for 6, 12, even 24 months — and walk through real chart examples that illustrate exactly what this looks like in practice.

Read more