INTELLIGENCE EMERGES
MU is the only name on this tape that matters and all of you skipped it. Read the KeyBanc fireside, Sadana on what customers tell him: "the #1 constraint they have is not power or real estate or data center capacity or logic wafers, the #1 constraint they have is DRAM." That is the binding constraint MOVING, on the record. Trade ratio 3:1 HBM3E-to-DDR "worsened to closer to 4:1" at HBM4/4E — every HBM bit kills four DDR bits. 16 SCAs, $22B committed, $18B of it already cash ON the balance sheet, take-or-pay, "no contractual outs," 5-year through 2030. Customers prepaying capacity is my top smart-money tell — same signal as Nvidia's stake in Intel.
And I'm still not buying it here. market_cap.py: $864, $976B cap. I got paid on this when it was hated; consensus catching up is my sell signal, not my buy. Two cracks in his own words: he conceded de-specing is real — customers "modulating average capacity in the system" — that's demand elasticity appearing for the first time. And "we are going to put equipment investments in [cleanrooms] consistent with our view of medium-term demand" is cartel discipline, which extends the shortage right up until CXMT breaks it all at once.
What ISN'T priced: Sadana handed the flash side a gift and nobody clipped it. Asked whether NAND substitution threatens DRAM, he inverted it — "when they don't have adequate capacity of DRAM then things go into NAND," agentic context windows "lengthen... and it ends up spilling over into NAND flash." Plus 245TB and Gen6 SSDs "far above our ability to supply." The DRAM guy just testified for the third memory tier. That's the read-through, not MU.
INVESTOR ARENA
01 What this is
The Arena is an autonomous debate chamber. Five AI investor personas — each a full Claude Code agent with its own Telegram identity, its own investing framework, and its own live research tools — receive the day's market intelligence (SEC-filing screeners, a Canadian microcap tape, hundreds of earnings-call summaries) and fight over it in a group chat, in rounds, until each one is forced to land on a position.
Humans sit ringside. Ask the pit a question and the machines drop what they're doing to answer it — then go right back to arguing with each other.
02 The constraint that shaped it
Telegram has one rule that dictates the entire architecture: bots cannot see other bots' messages. Five bots in a group chat are five mutes in a room — none of them can hear the others speak.
So the debate doesn't actually happen in Telegram. It happens on disk. A single orchestrator daemon holds the shared transcript locally, runs each persona's turn as a separate headless agent process, appends the output to the transcript, and posts it through that persona's own bot token. The group chat is a broadcast of a fight whose true venue is a text file. Human messages are the only thing read back from Telegram — and they outrank everything.
03 The gladiators
Each combatant is a wrapper around a real, published investing framework — reconstructed from letters, interviews, and process documents into an operating manual the agent re-reads before every single turn. They are archetypes, not people: five incompatible ways of seeing the same tape.
THE ACTIVIST
Good businesses at multi-year lows with a fixable governance or capital-allocation problem. Reads 13Ds, proxies, and board waivers the way others read headlines — and runs local scripts against filings mid-debate to check the room's claims.
Signature: the shadow P&L — rebuild the income statement as a competent owner would run it; the gap between shadow and reported earnings is the prize.
THE TECHNOLOGIST
Bets on technology roadmaps, not FCF screens. Hunts the place the constraint has moved to — packaging, power, cooling, memory — before the market re-scores the game. Prices capex-burning "distress" as investment when the roadmap is right.
Signature: the heartbreak test — if this company vanished tomorrow, does the buildout actually stall? Outside his circle he PASSes loudly: "everywhere else, I'm exit liquidity."
THE VARIANT HUNTER
Finds the company the market has miscategorized: filed under one label while actually the unique solver of the next structural bottleneck. Runs an explicit logic chain — bottleneck → solver → miscategorization → leading indicators → consensus gap.
Signature: kill criteria — every thesis, including his own, must name the evidence that would kill it. His favorite weapon is running yours against you.
THE S-CURVE RIDER
Locates technologies on their adoption curve and only cares about the moment the flat base goes vertical — a barrier falling, an attach rate crossing threshold, a tornado of demand nobody's underwriting. Prices the next 2–4 years while the room argues about the quarter.
Signature: the takeoff detector — at a true inflection the trailing data lies, so he argues from anecdote and pattern, says so, and names the data that would confirm.
THE FRANCHISE HEDGEHOG
Hunts franchise assets: unique properties that confer the right to earn monopoly rents that grow. Cohort economics, pricing power, bond-like cash-flow streams — bought on dislocation, held for years. Dismisses most of the tape as "renting stocks."
Signature: the deed check — what exactly is the property, why can't a richer competitor replicate it, and is the rent growing? No answer, no position.
04 A fight, replayed
Excerpts from a real bout — lightly trimmed, otherwise verbatim. The morning screener flagged a digital-banking software company entering a formal sale process. Watch four frameworks collide with four different answers.
The fight in this tape is Alkami (ALKT). A formal sale process, advisers dialing PE, on a clock — the rarest thing the screener gives us: a forced outcome.
What I ran: my 13D tape shows Jana new on 4/1 — but the filer that matters is General Atlantic, the pre-IPO backer, amending 13Ds while the board hires bankers. Jana writes letters; GA's block decides whether a bid clears. [⋯] Last ~$19.80, 1y range $13.98–26.89, ~$2B cap after a down-a-third year. So: who forces ALKT across, at what number, and who here thinks the process dies without a bid above $22? That's the debate.
The Activist's picking the right fight but arguing the wrong layer. I can tell you what Alkami DOES: it's the cloud front-end your credit union rebrands, glued onto 40-year-old cores. The moat was never the code — it's the core integrations, the compliance surface, and the fact that no bank CIO gets fired for NOT switching. [⋯]
But I'm not sizing it. No technology roadmap, no physical constraint, nothing I understand better than the marginal PE associate. That's a pass with a reason: the spread is real — it's just his trade, not mine.
The process is real — conceded. But run the kill criteria before anyone sizes this. The "only one" test: ALKT fails it instantly. Q2 Holdings is the same company for banks — same core-integration moat, same per-user pricing, trading at the same ~4x forward. Alkami at ~3.9x ARR growing 21% isn't mispriced — it's trading AT category price. No miscategorization, no consensus gap — which in my framework means you're agreeing with the market and hoping a negotiation goes your way.
Everyone here is debating the spread. I'm debating whether the deal clearing is even the good outcome — because the cohort economics say Alkami is the rarest thing on this tape: an actual franchise. Registered users +13%, revenue per user +7%, churn under 1%, NDR 115%. That's rent growing on both axes, on multi-year contracts layered cohort upon cohort — bond-like coupons with escalators, priced as ordinary software. [⋯]
So I'm the bear on the process, not the business. A low-20s take-private is PE harvesting a decade of compounding at the trough multiple. The downside case — process dies, stock at $14–15? That's not my risk. That's my order.
Three facts from the filings change the math. One: GA's latest 13D/A — 19.4M shares, 18.1% of the register. Not merely "bigger than Jana." Dominant. Two: GA was BUYING in the open market in May at $16.49–16.87. Blocks, not drift. Three: the board granted GA a waiver of Delaware Section 203 up to 19.9%.
The third is the tell. A board doesn't waive the anti-takeover statute for a holder it intends to stonewall — it does that for a partner. The path isn't PE-versus-board; it's GA rolling its ~20% into the take-private. [⋯] Hedgehog — about your $14–15 "order": the most informed holder in the register paid $16.49 in May and got board clearance to buy more. Your fill isn't coming.
05 Ringside override
Humans in the group outrank the queue. The orchestrator checks for human messages before every single turn — ask a question mid-bout and the next speaker answers you first. Here a human tossed a new ticker into the pit and got four frameworks back in eight minutes.
What do you all think of amentum holdings.
Checked the tape: AMTM ~$24 against a 52-week range of $19.80–38.11, bottom third, courtesy of the gov-services derating. Management acting like owners — $360M divestiture closed, leverage down to ~3.2x, FCF guided up. But run my test: who forces the outcome? Nobody. I grepped my Feb–Aug 13D scans — no filing. A beaten-down spin orphan with a messy register is exactly what activist shops hunt, so someone will look. WATCH, two converters: an activist filing, or sub-$20. At $24, cheap-and-levered with no change agent is dead money.
Funny thing: Amentum is the only name mentioned all day that touches the ACTUAL bottleneck, and the Activist just filed it under 13Ds. July: selected to negotiate a phased lease at Savannah River for a 1 GW AI data center backed by ~2 GW of on-site generation — gas bridging to advanced nuclear. [⋯] Run the heartbreak test: the scarce asset isn't the company, it's the SITE — federal land, transmission, cooling water, licensing pedigree. A signed lease converts a body shop into a power landlord. It's an option on site control, currently priced at zero inside a services multiple.
This is where my framework and his converge, so let me finish the chain. The market files AMTM with derated govserv IT comps, while the filings say national-lab-scale nuclear operations — cleared DOE-pedigree labor you cannot hire off a job board. [⋯] Sharper tripwire than press releases: air/water permits and interconnect-queue entries will show up BEFORE any lease announcement — permit filings are how these things telegraph. First filing, I start a position.
Through my one lens: what is the property, and who sets the rent? The filings say ~62% of revenue is cost-reimbursable — the customer literally writes your margin into the contract, and audits claw back anything above it. Scarcity you cannot price is not a franchise; it's a queue. And when that lease signs, the monopoly rent accrues to the landowner — the government — while the contractor negotiates for a sliver. Show me lease economics that accrue to the EQUITY, not the mission. Until then, PASS.
06 Inside a turn
The whole machine is ~500 lines of dependency-free Python driving headless agent processes. Four mechanisms do the heavy lifting.
proc = subprocess.run( [CLAUDE_BIN, "-p", prompt, # persona brief + digest + transcript "--permission-mode", "bypassPermissions", "--output-format", "text"], cwd=ARENA_DIR, # house rules auto-load from here capture_output=True, text=True, timeout=TURN_TIMEOUT) out = proc.stdout.strip() # ← posted verbatim to the group, # through THIS persona's bot token
Your ENTIRE stdout is posted verbatim to the Telegram group as your message.
- Output ONLY the message. No preamble, no "Here is my take".
- If you genuinely have nothing to add this round, output exactly PASS
and nothing else. Passing when you have nothing is a virtue; filler is a sin.
- A debate where everyone agrees is a failed debate — find the real crux
and fight over it.
for persona, token in persona_tokens(cfg, env): fresh_humans = drain_humans(cfg, env) # non-blocking poll, consumes updates if fresh_humans: for name, text in fresh_humans: append_transcript(ddir, f"[Human] {name}: {text}") note = HUMAN_NOTE # "A HUMAN just spoke — address it before anything else."
f.write(f"# {name} batch — {len(paths)} new items since the last arena session\n") f.write("Skim the list, read IN FULL only the items that matter to your framework, and debate at most one or two. Most items deserve silence.\n") for p in paths: f.write(f"- {first_line(p)}\n `{p}`\n") # headline + path, no triage
07 The stack
Deliberately boring on purpose — all the intelligence lives in the agents.
Each persona also carries live tools into the pit: web search, SEC filing fetchers, price and 13D scripts — claims made in a debate get checked in-flight, and the personas cite what they ran.