Tech buzz

Today's read

Catalyst hitFRVOCommercial operation · 83

Capacity fundedAMPXDoD award · 60 conviction

Technology ≠ stockNVTS37 conviction

Technology ≠ entryCOHU52 after another rally

New research slotTSSIAI capacity · prove conversion

Week of Oct 04

Evidence check

Adoption ahead Balanced Expectations ahead

Research prices · Oct 02 close · Tap a row for the thesis

Follow the checks

Capital signals

Binding commitments behind power, packaging, fiber, cooling and firm electricity.

01Fervo → grid33 MW COD · Sep 30Revenue under a contracted PPA

02Amazon → Generac$2.4B initial deliveriesBackup power in 2027–28

03Amazon → QualcommMulti-generation silicon dealAI inference + 1.6T optical links

04NVIDIA → SB Energy / OpenAI$105B capped guarantee4.25 GW Ohio AI campus

05Google → Fervo396 MW · 15-year PPACape Station delivery from 2028

06Flex → EPC Power$4.4B acquisition15+ GW deployed power conversion

07Anthropic → ERock470 MW order$1.7B contracted backlog

08NVIDIA → Amkor$1.5B prepaymentAdvanced packaging capacity

09NVIDIA → Coherent + Lumentum$2B eachOptical interconnect and photonics

010Meta → CorningUp to $6BU.S. fiber and connectivity

011Hyperscaler → Modine>$4B capacity agreementData-center cooling

012Google → FervoUp to 3 GW frameworkEnhanced geothermal

Read: the less-crowded opportunity is the physical conversion layer, not “AI” as a label.

Decision log

Weekly decision log

Scores change only when the evidence changes.

Oct 042026

2 stronger · 1 weaker on price · 5 unchanged

Fervo reaches commercial operation; federal funding strengthens Amprius manufacturing.

FRVO rises to 83 after its first 33 MW GeoBlock met the contracted output threshold, declared commercial operation one day early and began generating revenue. AMPX rises to 60 after a milestone-based DoD manufacturing agreement made up to $75M available, with $22M obligated and no company cost share; this funds capacity but does not prove customer demand. COHU falls to 52 on price only after another 9.7% rise without a durable HPC booking. The other five theses did not change adoption stage.

1ITRI842AMKR843FRVO834EROC785SANM726AMPX607COHU528NVTS37

Sep 272026

1 stronger · 2 weaker on price · 5 unchanged

Fervo reaches first power; Sanmina and Cohu outrun unchanged operating evidence.

FRVO rises to 78 after Cape Station synchronized to the grid and began selling electricity—the first utility-scale greenfield EGS project to reach first power. The 10.2% stock decline lowers expectations as adoption advances. SANM falls to 72 and COHU falls to 55: their 12.6% and 16.0% rallies arrived without new operating proof. These are stock-price downgrades, not technology downgrades. The other five theses did not change stage.

1ITRI842AMKR843EROC784FRVO785SANM726COHU557AMPX548NVTS37

Sep 202026

1 stronger · 7 unchanged

Amkor earns an evidence correction; Itron launches a product without yet advancing adoption.

AMKR rises to 84 because customer commitments already exceed Phase 1 Arizona capacity, supporting a 60,000-square-meter Phase 2 and roughly $12B total plan. The September 8 disclosure was missed in last week's snapshot, so this is labeled an evidence correction; execution risk also rises with construction stretching through 2029. ITRI's IEE Cloud AI launch improves the product set but has no disclosed paying deployment, leaving adoption and conviction unchanged. The other six theses did not reach a new stage. Amazon's Generac agreement is strong paid demand for backup power, but the stock rerated sharply.

1ITRI842AMKR843EROC784SANM775FRVO686COHU617AMPX548NVTS37

Sep 132026

2 weaker on price · 6 unchanged

No adoption-stage change; rallies narrow the margin of safety at Sanmina and Cohu.

SANM fell to 77 after an 8.8% weekly rise without a matching operating disclosure. COHU fell to 61 after a 12.5% rise while durable HPC bookings remained unproven. These were stock-price downgrades, not technology downgrades. FRVO fell 14.2%, improving entry after its Google PPA, but Q4 test power remained unhit.

1ITRI842AMKR803EROC784SANM775FRVO686COHU617AMPX548NVTS37

Sep 062026

1 stronger · 1 weaker · 6 unchanged

Fervo converts hyperscaler interest into a binding PPA; 800V power validates faster than Navitas.

FRVO rises to 68 after a guaranteed, 15-year Google PPA for 396 MW, despite a 23% weekly rally. NVTS falls to 37: Flex is paying $4.4B for an 800V-capable platform with 15+ GW deployed while Navitas still lacks a named production win. AMPX's SiCore500 is promising technology, but not paid adoption. The other five theses produced no material operating change.

1ITRI842SANM803AMKR804EROC785FRVO686COHU657AMPX548NVTS37

Aug 302026

1 weaker · 7 unchanged

Navitas buys more technology, not adoption; Fervo improves risk oversight without reaching first power.

NVTS falls to 40 because the Claros acquisition adds cash-and-stock consideration, integration work and a 2028–29 payoff while its named paid-production catalyst remains unmet. FRVO's independent seismicity review is constructive risk evidence, not an adoption-stage change. The other six companies produced no material operating update; lower prices alone did not trigger score changes.

1ITRI842SANM803AMKR804EROC785COHU656FRVO627AMPX548NVTS40

Aug 232026

1 evidence correction · 1 price upgrade

Fervo's construction stage was understated; AMKR improves as a stock, not as a technology thesis.

FRVO rises to 62 after correcting the missed Q2 release: two GeoBlocks are mechanically complete, commissioning is underway and Q4 test power is the next hard checkpoint. AMKR rises to 80 because a 14.8% decline improved the entry without adverse news. The other six theses are unchanged; broad selloffs lowered expectations but did not create new adoption proof.

1ITRI842SANM803AMKR804EROC785COHU656FRVO627AMPX548NVTS43

Aug 162026

1 upgrade · 2 repricings

ERock earns the only evidence upgrade; price moves alter the asymmetry elsewhere.

EROC's Anthropic order and manufacturing progress outweigh its 58% rally, raising conviction to 78. COHU falls to 65 because a 12% move arrived without new operating proof. FRVO becomes modestly more attractive after a 24% decline, but remains pre-revenue and firmly in the watch tier.

1ITRI842SANM803EROC784AMKR775COHU656AMPX547FRVO548NVTS43

Aug 092026

Baseline established

Physical AI infrastructure leads; narrative-heavy moonshots stay on watch.

Highest confidence begins with ITRI, SANM and AMKR. EROC offers more upside variance with project-execution risk. AMPX, FRVO and NVTS require proof that operating results can catch expectations.

ITRI84SANM80AMKR77EROC72COHU68AMPX54FRVO51NVTS45

New-candidate scan

TSSI enters research; ACCV's discounted IPO still carries sponsor and dilution risk

TSSI

AI-integration capacity is real; concentration and valuation are not hidden

Q2 systems-integration revenue grew 46% and facilities-management revenue grew 84%. A $17M next-generation AI data-center investment is expected to begin converting into integration revenue in Q3. At $10.24 and roughly $289M of equity value, the attention gap is interesting—but about 79× trailing earnings, cancelable contracts and customer concentration block promotion. Falsifier: Q3 fails to convert the investment into integration revenue and gross profit.

Evidence ↗

BTDR

Real prepayments; larger expansion still needs customers

Bitdeer's 9.5 MW A102 site is sold out under five-year commitments exceeding $800M, with customer prepayments expected to cover more than 50% of capex. Against a Sep. 25 equity value near $1.58B, that deserves research. But A201 and A202—86.8 MW combined—had no offtake commitments as of September 14. Falsifier: A102 misses Q1 2027 energization or prepayments fail to limit capital strain.

Evidence ↗

ACCV

IPO priced down; sponsor exit and dilution still dominate

Accelevation priced at $18, below the $20–24 range, and closed October 2 at $17.72. The backlog is real, but only 10M of 30M offered shares came from the company; selling stockholders supplied 20M, with another 4.5M-share option. The filing implies about $3.94B of equity value and substantial tangible-book dilution. A discount to the offer does not yet create an attention gap. Falsifier: backlog fails to convert into revenue and cash after the IPO.

Evidence ↗

FPS

Exceptional paid adoption; valuation closes the gap

Forgent reported $1.50B of quarterly bookings, 3.3× book-to-bill and $3.0B backlog, with Q4 adjusted EBITDA margin of 24.4%. A $35M Tijuana expansion will lift total revenue capacity toward $5.8B. At $39.44 and roughly 304.5M Class A and B shares, the implied equity value is about $12B: excellent operating evidence, but not an underfollowed bargain.

Evidence ↗

FCEL

Paid reservation; economics and dilution block promotion

FuelCell Energy has a paid 75 MW data-center capacity reservation, a firm 30 MW initial Fit Energy phase and $1.3B of committed backlog. The remaining $2.35B awarded backlog is optional, Q3 gross loss was $24.5M on $33.0M revenue, and shares outstanding rose from 46.1M to 80.0M in nine months. At a Sep. 11 equity value near $1.12B, this is high-signal research—not an attractive stock yet.

Evidence ↗

FLNC

Backlog-price gap; execution blocks promotion

Fluence has $6.4B of backlog and $850M of data-center business against a Sep. 11 equity value near $1.37B. But Q3 gross margin fell to 5.1%, net loss was $44M, and $400M of deliveries slipped because new manufacturing facilities lagged. Real adoption; not yet an investable execution profile.

Evidence ↗

DUOT

Paid adoption; dilution is the catch

Five-year hosting agreements total 65 MW and more than $611M, while 25 MW is contracted for 2026. At a Sep. 11 equity value near $294M the attention gap persists—but prior share growth plus pre-funded warrants keep dilution central. Research slot only.

Evidence ↗

POWL

Price cooled; still not undiscovered

At $182.50, the operating evidence is unchanged: $934M of quarterly orders, 3.0× book-to-bill and $2.4B backlog. At about 35× trailing earnings and a $6.7B equity value, adoption remains better than the attention gap.

Evidence ↗

GNRC

Amazon validates adoption; the stock catches up

A long-term Amazon agreement covers about $2.4B of expected 2027–28 deliveries and a warrant whose later tranches vest as purchases rise toward $8B. The deal is decisive paid-adoption proof, but the shares jumped and closed at $207.44, about a $12.4B equity value and 48× trailing earnings. The technology and demand strengthened more than the stock opportunity.

Evidence ↗

FLEX

800V control group; scale is already expensive

Flex agreed to pay $4.4B for EPC Power, which reports more than 15 GW deployed and expects about $800M of 2026 revenue. The deal validates 800V power conversion while raising integration, debt/equity financing and valuation risk; at roughly 41× trailing earnings, this is a competitive benchmark, not an underfollowed bargain.

Evidence ↗

Reference shelf

Research & method

Kept here when you need it—not in the way of the daily scan.

AI investing researchLive, forward-paper and backtests · updated Oct 04+

Autonomous Market Intelligence

Forward-collected paper strategy · not real-money live

Daily Russell 1000 predictions have been generated out of sample since April 2025 with autonomous web research. The reported top-20 long portfolio produced positive factor-adjusted returns, but the signal was concentrated in the very top names and remains a research portfolio rather than deployed capital.

Inspect the project ↗

KTD-Fin

Leakage-controlled historical benchmark · not live

A masked CSI 300 evaluation with Barra attribution found that apparent gains were largely market and style exposure, with limited persistent stock-selection alpha. It is a useful warning that fluent rationales do not establish security-selection skill.

Inspect the project ↗

NextFund

Live-market evaluation infrastructure · no verified returns

Time-consistent market access and persistent logs connect each observation, justification and trade across U.S., Hong Kong and China A-share markets. That improves auditability, but the paper reports a platform demonstration—not evidence of excess returns.

Inspect the project ↗

DX production fleets

Forward live · real-money + live-price paper fleets

A six-month record covers 3,505 real-ETH agents and 500–599 mostly paper perpetuals agents. Neither fleet showed a directional edge: only 16.2% of the real-money vaults finished profitable, while the paper fleet lost money and trailed matched retail. Controls and order-path mechanics mattered more than prompt prose.

Inspect the project ↗

FinSkillBench

Point-in-time task benchmark · not live trading

Across 9 models and 2,603 investment-management episodes, curated procedural skills lifted mean scores from 0.366 to 0.528; agents writing their own skills gained little. Reliable process scaffolding mattered more than improvised prompt complexity.

Inspect the project ↗

LLM Trading Lab

Forward-only, public microcap experiment

Excellent transparency: daily logs, trades, benchmarks and risk rules. Early gains are interesting, but one tiny portfolio is not statistical proof.

Inspect the project ↗

1rok

Forward test · public paper accounts

Seven models receive the same tools, prompts and $100K paper accounts on a weekly clock. The open harness is useful because research and order execution are separate; its leaderboard is evidence about one live paper protocol, not real-money alpha.

Inspect the project ↗

InvestLogicBench2026

Forward test · simulated capital

Across seven weeks, most agents trailed SPY; the best general reasoning scores did not predict returns. The durable lesson is to score event selection and thesis continuity, not model prestige.

Inspect the project ↗

AQuA

Causal walk-forward backtest · not live

Self-improving research agents carried empirical beliefs across runs inside sealed evaluation sandboxes. Promising process evidence, but still historical and not a forward portfolio result.

Inspect the project ↗

010

CLQT benchmark

Backtest + four-week live paper track

The best Sharpe ratio was not the most capable agent. Models often traded in the direction of their analysis but sized positions inconsistently with their own reasoning.

Inspect the project ↗

011

FIDES

Out-of-sample backtest protocol · not live

FIDES checks whether an agent's prose, code and realized track record describe the same strategy. Only 2 of 40 generated strategies beat buy-and-hold, and claimed outperformance was badly calibrated.

Inspect the project ↗

012

builderr Trading Agent

Forward test · shared paper sandbox

Its July 7–September 4 round used common fills, fixed risk limits and a public agent brief. The repository describes the protocol but does not yet publish a final leaderboard, so no performance claim is credited.

Inspect the project ↗

013

Vibe-Trading

Open-source prompts and tooling · no verified returns

A useful library of research, backtest and bull/bear committee prompts with explicit paper/live connector controls. It is implementation evidence—not evidence that the generated strategies outperform.

Inspect the project ↗

New findingNo new verified real-money alpha surfaced. Autonomous Market Intelligence offers encouraging forward-collected, out-of-sample paper evidence, while KTD-Fin's masked historical benchmark attributes much apparent performance to market and style exposure. The labels matter: a forward-collected paper portfolio is stronger than a conventional backtest, but it is still not live capital or proof of durable alpha.

Weekly research protocol4 evidence rules + prompt+

01
Prove adoption
Rank prepayments, binding orders and paid deployments above pilots and MOUs.
02
Price expectations
Measure valuation, attention, financing needs and what success the stock assumes.
03
Force a bear case
Name one observable catalyst and one falsifier before assigning conviction.
04
Compare last week
Change the score only when evidence changes. “No change” is valid.

WEEKLY_REVIEW.md

OBJECTIVE
Find adoption–expectations gaps.

FOR EACH TICKER
1. Point-in-time evidence + source
2. Theme revenue materiality
3. Adoption score (0–5)
4. Expectations score (0–5)
5. Catalyst and falsifier
6. Conviction delta vs prior week
7. Explicit no-action reasons

RULES
Primary sources first. No lookahead.
Mark missing data. Never invent a
price target or weekly change.

Sources & standardsPrimary evidence first+

Company filings, earnings releases and binding commercial announcements anchor the stock work. Public repositories and papers anchor the AI-agent research. Watchlist prices are research snapshots; the top tape supplies current market context.

AI Hedge Fund TradingAgents ATLAS TrustTrade Expert Teams CLQT live benchmark InvestLogicBench AQuA walk-forward test 1rok forward paper benchmark FIDES concordance protocol builderr forward-paper protocol Vibe-Trading prompts DX production-fleet record NextFund live-market evaluation Autonomous Market Intelligence KTD-Fin attribution benchmark NVIDIA / OpenAI commitment