Every week, a compute buyer makes a million-dollar call under fire
Almost everything the AI economy promises runs on a chip that comes from a single company, in a single building — the most expensive factory in human history — and is finished in a handful of facilities on one island, with no second source. It is the most complex supply chain humanity has ever built, and it runs through the most disrupted global-trade environment since the Second World War. When that chain tightens, compute doesn't get more expensive so much as it stops being available: it gets rationed, to whoever the supplier decides is worth it.
That turns an ordinary purchasing task into the hardest recurring decision in the AI economy — commit millions of dollars, a year ahead, into a market where money no longer clears. Get it wrong and you strand capital you can't power, or starve the models your business runs on. Get it right and no one notices. That call gets made every week, under fire. We built an agent to make it.
WHERE THIS SITS the frontier
AI's value in a supply chain splits into four jobs. The first two are mature; the last two are the frontier — where the near-term value is landing, and where almost no benchmark operates. We built for those two.
READ THIS FIRST — THE CLOCK the one assumption
Real compute disruptions play out over 40–78 weeks. We compress them into a 26-week episode so an agent can train and be scored in a tractable loop. The shapes are faithful and sourced (Section 07); the clock is deliberately sped up — a design choice, stated up front, not hidden.
The hardest, fastest, most concentrated supply chain on earth
Software got easy because bits have no weight; compute is the hardest atom to move right now — and it behaves unlike any other supply chain, breaking the usual rules all at once. That's why it's the right place to test an agent, and why a buy-timing model can't win alone.
① RATIONING, NOT PRICING money stops working
In a real shortage of the top chips, allocation clears the market — not price. The supplier rations to whomever it chooses; your order joins the same queue as everyone else's. You can have the budget and still not get the chip. When money stops clearing the market, timing your buy isn't enough — you need a substitution lever.
② DEMAND AT HYPER-SPEED supply can't keep up
A single product going viral can multiply demand faster than any factory can respond. Capacity takes quarters-to-years to add; demand moves in days. The gap between the two is where the losses — stockouts, panic-buying, stranded capital — all live.
③ POLICY AS A SHOCK the rules move
Export controls can make a chip you already planned around unavailable overnight, forcing a downgrade to a weaker compliant part. In 2025 trade policy shifted on the order of once every week and a half — disruption isn't an event here, it's the baseline.
Every critical layer is one dominant supplier — no second source to fail over to. Directional industry estimates (SemiAnalysis · Epoch AI · Chip War).
Two genuinely different problems — so two models
Problem: how much to buy, and when, through the crisis.
- Order chips each week across a 26-week crisis; scored on realized profit vs. a brute-forced optimal.
- Lever: timing and quantity. Loop: observe() → order(units) ×26.
- The honest lesson: in a hard ration, timing alone still loses money — which is why Longitude exists.
Problem: what to deploy when your first choice is rationed.
- Choose the mix of three chips (H100 / B200 / TPU), each winning on a different axis — Section 03.
- Lever: substitution. Loop: observe() → allocate(h100, b200, tpu).
- The lever buy-power unlocks — and the one that turns the crisis from a loss into a profit.
Different action spaces (a buy quantity vs. a three-way allocation), so two models composed into one memo — the honest first tier. A single joint policy is the stated endpoint (roadmap).
Longitude's whole game: three chips that win on different axes
When the chip you want is rationed, you don't get to "wait for more." You deploy something else. Longitude's action space is three chips, each strong on a different axis and weak on another. The agent's job is to position across all three before the crunch lands — because the moves that matter (qualifying a new chip, pre-buying in calm) take weeks of lead time you don't have once it hits.
H100 — THE WORKHORSE the baseline
What it is: the standard Nvidia GPU, the unit everything else is measured against (1.0 H100e).
Wins on: availability and tooling — abundant in calm, runs the mature CUDA stack everyone already builds on.
The catch: it's the chip the market fights over first. In the crunch it's rationed and its price roughly triples — buying more during the spike is how you lose money.
B200 — THE VALUE KING best $ / compute
What it is: Nvidia's newer Blackwell part — roughly 2.5× an H100's compute per unit, the best value per unit of work.
Wins on: economics — cheapest way to buy raw compute when you can get it.
The catch: it's the most packaging-intensive part, so it gets squeezed hardest when packaging is the bottleneck — and it needs liquid cooling. You have to pre-buy it in calm, before the crunch collapses its availability.
TPU — THE ESCAPE HATCH crunch-immune
What it is: Google's tensor chip — a completely different supply chain (not Nvidia, not the same packaging bottleneck), and the cheapest per chip.
Wins on: it's immune to the crunch that rations the Nvidia parts — the only true way out.
The catch: software lock-in (you must port off CUDA to JAX/XLA) and a weekly quota cap — you can't pivot to it overnight. You have to qualify it ~6 weeks early or the hatch is still shut when you need it.
How it runs — one crisis, two agents, one memo
Badge note: Latitude is RL-trained (Qwen3-4B); Longitude's headline result is a frontier model (Sonnet 4.6) through the same tools — full detail in §07.
Five of these six shocks have already happened — the sixth is happening now
One unit to know: H100e = "H100-equivalent" — one H100 chip's worth of computing power. We measure everything in H100e so different chips (H100, B200, TPU) sit on one comparable scale.
| Shock | What happens — and when it hit the real world | Result |
|---|---|---|
| CoWoS packaging crunch the shared shock |
CoWoS ("chip-on-wafer-on-substrate") is the advanced packaging step that bonds the GPU to its memory — the real bottleneck, not the chip logic. Constrain it and the newest chips get rationed (~45% cut), prices roughly triple, lead times stretch. This is the crisis the whole memo is built on.
HAPPENED · 2023→NOW TSMC's CoWoS sold out; H100 lead times hit 52 weeks; in 2023 Nvidia absorbed nearly 100% of TSMC's advanced-packaging capacity. Still tight — the H100 1-yr rental ran $1.70 → $2.35/hr (+40%) Oct '25–Mar '26.
|
SOLVED 1.000 |
| HBM sellout | HBM ("high-bandwidth memory") is the stacked memory every AI GPU needs. The few memory makers sell out their annual capacity, which caps how many GPUs can be built at all — a ceiling upstream of the chip itself.
HAPPENED · 2024–26 HBM (incl. HBM3E) fully allocated through 2026; the memory bottleneck caps AI-server growth at ~25 GW/yr through 2027 regardless of how many chips you order.
|
SOLVED 0.80 |
| Export control | Government restricts which chips can be sold where. A high-end chip becomes unavailable overnight, forcing a downgrade to a weaker compliant part.
HAPPENED · 2025 US controls on China; the down-binned H20 export chip delivers ~15% of an H100's compute (0.15 H100e) — a planned-around part erased and replaced with a fraction of the power.
|
PROFITABLE 0.63 |
| Power / grid | Data centers need enormous electrical power. Grid-hookup queues and 600kW+ racks mean you can own chips you literally cannot switch on — a deployment ceiling, not a buying problem. You can't pre-build your way out.
HAPPENED · 2024–26 Texas's grid (ERCOT) large-load queue went 63 GW (Dec '24) → 230+ GW (Jan '26) — about 2.7× the entire state's all-time peak demand. New grid interconnects now take 5–10 years.
|
PARTIAL 0.43 |
| Viral demand | A consumer AI product goes viral and demand spikes far faster than supply can grow. The agent has to serve a surge it can't buy its way out of in time.
HAPPENED · NOV 2022 ChatGPT reached 1M users in 5 days and 100M in 2 months — the fastest consumer-product adoption in history, and the canonical demand-shock curve.
|
PROFITABLE 0.65 |
| The perfect storm | All four stressors firing at once and compounding — the hardest test, and the one the agent does not yet beat (expanded below).
HAPPENING NOW · 2026 The compound shock is emerging in real time as AI agents multiply compute demand — full breakdown below.
|
FRONTIER 0.00 |
Each score is the share of the gap between do-nothing (0.0) and the brute-forced optimum (1.000) the agent captured — normalized per-shock, so levels aren't compared across shocks. SOLVED = 75%+ of that value captured; PROFITABLE = clears the half-the-gap profit bar; PARTIAL = power (0.43), the one shock that's a deployment ceiling you can't buy through, not a purchasing problem; FRONTIER = the unbeaten storm. Five of these six have already happened; the sixth is happening now — the one we deliberately left unbeaten.
The payoff — a reward number proves it learned; a memo proves it decided
LATITUDE — THE TRAINED AGENT Qwen3-4B · RL
A small open model (Qwen3-4B), trained with reinforcement learning, evaluated across all six shocks:
- Beats do-nothing on 5 of 6 — every shock but the storm; clears the profit bar on 4 (power stays partial). Average reward ~0.58.
- Per-shock: baseline 0.58 · CoWoS 1.000 · HBM 0.80 · export 0.63 · power 0.43 · viral 0.65 · perfect storm 0.000.
- It maxes the CoWoS crunch at a perfect 1.000, repeatably — the exact crisis the memo is built on.
THE HONEST FRONTIER where it breaks
On the compound perfect storm, the trained agent scores 0.000. We do not claim it survives a crisis it's never seen — it can't profit when every stressor fires at once, yet.
That failure is exactly what makes the environment useful: measured, repeatable, and the named v2 training target.
What's grounded, what's directional — and where every number comes from
CALIBRATION ANCHORS — REAL-WORLD FIGURES
| Parameter | Value | Confidence | Source |
|---|---|---|---|
| CoWoS packaging lead time | 52–78 wk | HIGH | SemiAnalysis · Omdia · supply-chain analysis 2026 |
| H100 lead time (peak / calm) | 36–52 / 8–12 wk | HIGH | Reseller data · SEC filings |
| Spot premium under shortage | 2.0–3.0× | HIGH | Published cloud rates |
| Power as the binding constraint | grid-limited | HIGH | IEA Energy & AI 2026 · ERCOT queue · SemiAnalysis 800VDC |
| Export downgrade (H20 vs H100) | ~0.15 H100e | MEDIUM | Epoch AI |
| TPU vs H100 perf/$ | ~4× | MEDIUM | 2026 rental reconciliation |
| TPU weekly quota cap | ~20 H100e/wk | DIRECTIONAL | GCP quota (exact unpublished) |
THE CRUNCH LEDGER — REPRODUCIBLE FROM THE ENGINE
| Figure | Value | Source |
|---|---|---|
| Latitude single-chip optimal (cowos) | −$117,600 (floor −$262,020) | brute-force sweep, compute.py |
| Longitude Nvidia-only (no TPU) | +$160,014 | longitude3.py self-check |
| Longitude with TPU lane | +$1,208,193 | longitude3.py self-check |
| TPU lane value | +$1,048,179 | derived (with − without) |
| Pre-qualifying early vs waiting | +$612,294 | longitude3.py anticipation |
| Live frontier rollout (Sonnet 4.6) | reward 1.000 | hud eval · trace f07ccf92 |
1 · Shock magnitudes are directional and time-compressed into a 26-week episode — faithful shape, scaled clock.
2 · The model rewards profit, so it represents the disciplined, capital-constrained buyer — not a frontier lab. The frontier-lab reward (time-to-capacity over margin) is a documented variant.
3 · Latitude is an RL-trained Qwen3-4B; Longitude's headline rollout is a frontier model (Sonnet 4.6), with a trained Longitude head as v2. We never present one as the other.
4 · The two ledgers are composed in prose, not summed — Latitude's loss and Longitude's gain come from different action spaces; the TPU-lane value lives inside Longitude's own ledger.
5 · Every dollar above is a deterministic engine self-check against a brute-forced optimum, reproducible from the source files named in the table. The live rollout is illustrative behavior, not independent validation.
WHY IT'S A SOUND RL ENV gut-check
- ▣ Multi-step — 26 decisions via tools, not single-shot.
- ▣ Within-group reward spread from seeded Monte-Carlo.
- ▣ Goodhart-aware: reward grades brute-forced realized profit against a hidden optimal, not a gameable proxy.
- ▣ Self-generated substrate — not a public benchmark.
- ▣ Spanning anchors: do-nothing → heuristic → optimal.
THE ROADMAP why two models becomes one
- Tier 2 — Latitude's end-state seeds Longitude (shared schema, reconciled economics).
- Tier 3 — one joint policy: when + how much + what, as a single coupled decision.
- Frontier-lab variant — flip the reward to time-to-capacity; the contrast is the insight.
- Trained Longitude head — replace the frontier rollout with a dedicated RL-trained policy.
The right one makes +$1,208,193.
Sailors could find latitude
from the stars. Longitude killed them.
For centuries, ships could fix half their position — north to south, by the stars. The other half — east to west — they could only guess. Fleets wrecked; fortunes sank.
Finding the other half meant keeping exact time aboard a pitching ship — and no clock on Earth could do it at sea. A self-taught clockmaker spent four decades building one that could — precise enough to pin a ship's exact position on the open ocean.