Grid & Gradient — Prototype Scope and Game Design

Status: Research-backed proposal · Date: 2026-11-06 · First mission: The Renewal

The player promise

Build a company whose models are useful enough to sell, cheap enough to serve, and reliable enough to keep customers. Your first two quarters left you with a small lab, a dense model fine-tuned from a licensed checkpoint, an elastic cloud deployment and $1.20M in cash. A rival has just undercut you. A renewal gets you to scale; a bad deal leaves you paying for racks with no buyers.

The intended player is a strategy or management player curious about the AI business. They can start without knowing what a token or a megawatt is. By the end they should be able to explain why the cheapest listed token, cheapest successful job, and best capacity deal can all be different choices.

Why this game shape

Approach What it buys Tradeoff
Real-time campus builder The strongest sense of watching a site come to life Construction, pathfinding and time pressure can crowd out the business decisions
Executive simulator with cards and spreadsheets Fastest to prototype and easy to audit Sites feel interchangeable; building and competition have little presence
Regional board with quarterly turns Physical infrastructure has a place, tech has a timeline, and a turn gives time to read the numbers Fewer buildable pieces and no rack-level simulation in the first mission

Choose the regional board. A turn starts with a rival move and a customer deadline, then lets the player change the plan and see a cash forecast before committing. Place a lab or compute hall where the fiber, power and customer links matter. Small animated flows can show where training jobs and API traffic go. This keeps the feeling of building something that persists across turns.

What it looks like

 GRID & GRADIENT     LATTICE LAB        Q3 / 8    CASH $1.20M
 ┌───────────────────── Regional board ──────────────────┬─────────────────────────┐
 │  RELAY RIDGE    ◇ Reserved cluster · ready Q4          │ THE RENEWAL             │
 │           ····· fiber ·····                            │ Sable cut its quote.    │
 │  HARBOR   ▣ Lab  ──  ▧ Meridian Cloud  ──  ▥ Client   │ 40M requests / quarter  │
 │                           │                            │ ≤ $0.006 / request      │
 │       ~ river ~           │ new fiber                  │ ≥ 86% task pass         │
 │  RIVERBEND    ▤ Colo pad ─ ⚡ Grid slot · ready Q5      │ ≤ 800 ms p95 first token│
 ├───────────────────────────┴────────────────────────────┴─────────────────────────┤
 │ Research  ·  Deals  ·  Ledger        QUARTER FORECAST               [Advance Q3] │
 │ Base + prefix cache on elastic cloud: $240k revenue · $168k serving · −$88k cash │
 └──────────────────────────────────────────────────────────────────────────────────┘

Aim for a 2D atlas seen from above with a slight isometric tilt. A river and power corridor anchor the map; the lab, substations and server halls are recognizable silhouettes. Keep the board quieter than the decisions. A lit solid link means usable capacity, a dashed link means signed or under construction, and a broken link marks an outage or delay. Each also has a text label. Select a site to see three useful facts first: what comes online when, what it costs before you serve a request, and what could stop it. The market panel shows the rival and customer; a compact ledger shows the consequence of the selected move. Research, contracts and source notes open in place. Use the same controls by keyboard, with a list view of the map's decisions for small screens.

The turn

  1. Read the market. The buyer compares your quote and service with a visible alternative. Inspect the workload and the rival's latest move.
  2. Commit two scarce teams. One infrastructure action and one lab action per quarter. Queue a site, sign a block, or keep cloud capacity elastic. Spend the lab action on serving, model work or evaluation. Pricing and selecting a cleared deployment are separate controls.
  3. Forecast and negotiate. Compare the proposed quote, quality, latency, serving expense, fixed obligations and estimated closing cash. See the difference between a binding purchase and a capacity offer. Read the downside at lower demand before signing.
  4. Advance the quarter. Apply delivered capacity, bill fixed commitments even if idle, sell only to contracts whose gates pass, and resolve any announced rival or supply change. Record who moved, what changed, and why the company won or lost that revenue. Revisit the last decision with a different choice.

In a longer campaign, the eras would be Lab → Product → Scale → Frontier. Product adds distribution and repeat customers; Scale adds multiple contracts, redundancy, energy and finance; Frontier adds large research programs and accelerator supply. Progress unlocks options rather than making ownership or frontier training mandatory. A specialized supplier, a broad API lab, and an independent infrastructure owner could each get a different objective. The first mission plays only the transition from Product to Scale.

First mission: The Renewal

Six turns, Q3–Q8, about 15–20 minutes as a playtest target. The opening board is fixed. The buyer is an enterprise coding assistant with a steady, reusable prefix and a published acceptance workload. The tutorial presents the Q1–Q2 history as a short briefing. All figures below are fictional, tuned for comparison. The board labels amounts as simulated.

Item Opening rule
Buyer 40M billable requests per quarter Q3–Q6; expansion to 65M in Q7–Q8 if the lab has served the renewal every quarter through Q6. This demand and trigger are visible from Q3. A missed quarter forfeits the expansion.
Tariff Old list: $2/M input and $8/M output. Renewal ceiling: $1.50/M input and $6/M output, equivalent to $0.006 for the fixture's 2,000 input and 500 output tokens. Quoting less gives no extra volume from this buyer. This negotiated tariff does not meter cache reads separately.
Gates At least 86% task pass on the protected workload, p95 time to first token (TTFT) no more than 800 ms at the planned load, and a cleared release. The lab forecasts against tested profiles. It cannot bill the renewal if a gate fails or if it cannot serve the full volume.
Sable Models The efficient rival's Q3 quote is $0.006 per request on the fixed mix. It meets the buyer's gates; the buyer favors the incumbent on a tie if it qualifies. The rival gives the player a credible outside option, with an announced repricing/downside scenario for replay.
Lattice Lab $1.20M starting cash; $160k fixed lab operations per quarter. Start with Base at 90% held-out pass, cleared. No revenue other than this buyer in the paper baseline; an optional fallback market belongs in later design.
Prefix cache One lab action, active in the current quarter; already included staff cost. Changes Base elastic serving cost from $0.0054 to $0.0042 per attempt for this eligible workload. Base retains its held-out quality. Cold one-off traffic gets neither this saving nor the warm-profile latency.
Compact One $55k lab action to distill a small dense candidate, results next quarter: 84%, below the gate. One subsequent $35k lab action for licensed data and eval work, next-quarter result: 88%, cleared on the v0 fixture. Serving Compact on elastic cloud is $0.0026 per attempt without assuming a cache saving. Base can serve the renewal during both projects.

Capacity offers

Option Earliest service if chosen Q3 Cash cost Workload profile
Meridian Cloud, elastic Q3 Per attempt only, at Base or Compact rates above Cap 65M/quarter in this scenario. Tested p95 at 40M/65M: Base cold 750/790 ms, Base cached 690/750 ms, Compact 540/620 ms.
Relay Ridge, neocloud reservation Q4 $120k each quarter for three quarters, including idle quarters, plus $0.0010 per Base attempt Cap 40M/quarter and p95 540 ms for the fixture. Over the cap, plan a second route; model mixed p95 only with a measured mix. No mixed route in the first mission.
Riverbend, colo pod Q5 $220k upfront; $64k fixed per quarter once live plus $0.0012 per Base attempt Cap 80M/quarter with p95 510/590 ms at 40M/65M. Click the site to see the waiting grid slot and fiber dependency. Bridge with the cloud until delivery. The variable rate includes the fixture's power and operating cost.

Use the premeasured fixture values for these packages; the scenario does not derive real latency from a price or average regional ping. Only tested combinations are selectable in the first mission. Relay and Riverbend host Base in v0. A signed Relay commitment continues to accrue even if traffic is assigned elsewhere. The colo's grid slot and fiber handoff are distinct; the later delay variant moves the active date and keeps the bridge bill.

One decision on paper

At 40M requests, the new customer quote earns $240k this quarter. Elastic Base serving costs $216k with a cold profile, leaving $24k of serving contribution before $160k fixed lab costs. Prefix caching cuts serving expense to $168k. After the same fixed costs, cash still falls $88k, which the CFO flags before you sign. A 90% acceptance rate yields 36M successful jobs for comparison: $168k ÷ 36M ≈ $0.00467 per successful job. The tariff bills attempts. A successful job passes the task and finishes within the 20-second deadline. The listed profiles assume all served jobs finish in time, so deadline misses do not change the count in this slice. TTFT is a separate contract gate. An unqualified candidate gets no renewal revenue even if its internal cost looks low.

At 40M attempts, Relay's $160k serving bill is $8k below cached elastic. At 20M, its $140k is $56k above cached elastic. The crossover on this particular workload is 37.5M attempts/quarter, before any one-time migration, and the service cap is 40M. The same reservation can be helpful against a latency deadline and a bad bet against a volatile customer. The forecast uses the entire minimum term rather than comparing the first invoices.

Success and meaningful paths

Win: never fall below zero cash, qualify for the expansion, and end Q7 and Q8 with the renewal served and positive operating cash before discretionary investment in both quarters. The board also retains evidence of release clearance. A missed renewal, unpaid obligation or failed liquidity gate gets a causal debrief. The target is a profitable service with a viable company behind it.

These paths show arithmetic feasibility for the fixed seed. Balance comes from uncertainty: a later site or canceled expansion can strand capital, while a harder workload can remove Compact's quality margin. For the first replay, try Second Cut: in Q5 Sable publicly announces a channel offer effective Q7 at $1.25/M input and $5/M output ($0.005/request for this mix). At 65M, matching it leaves Compact at −$4k and Riverbend at +$23k per quarter after fixed lab costs. There is time to order Riverbend in Q5 for Q7. Keep the 20M forecast as a separate sensitivity check. The player should be able to explain when each path is attractive.

The research tree after this slice

Branch What to expose when it becomes a decision
Models and data Dense vs sparse MoE, total and active weights, memory/fabric fit, a data source's rights and coverage, pretraining vs specialization, post-training and reasoning effort. A good coding model need not be a good multilingual model.
Serving Prefill and decode profiles, cache eligibility and expiry, continuous batching vs asynchronous batch products, quantization quality checks, route/escalate with retries, and capacity headroom at peak.
Capacity and energy Spot/on-demand vs firm blocks, reserved start dates and minimum terms, ownership and depreciation, second fault domains, grid connection and cooling, offtake and curtailment. Show MW for instantaneous power and MWh for electricity bought or consumed.
Distribution and capital Direct API vs cloud marketplace, enterprise contracts vs consumer subscriptions, dilution, debt service, credits, channel fees and partner dependence. A large order can create a cash problem even when its contribution is positive.
Governance and rivals Safety evaluations and staged release, audit and rollback, public scrutiny, customer eligibility, coordinated supply and reactive rivals. The player sees the reasons, not a hidden morality meter.

Prototype boundary and learning check

Start with a browser-sized, single-player, deterministic 2D board. Use a small simulation independent of presentation, explicit input units, a frozen scenario version, forecast/settle arithmetic from the same rules, and a quarter log. Draw initial buildings and links with simple original shapes. No live pricing feed, real model calls, accounts, multiplayer, full 4X campaign or rack logistics in the slice. No technical dependency needs to be chosen to test the first decision on paper.

The player learns from a choice and its receipt: click the delta to see how many jobs were billed, which fixed obligation got paid, and which quality or service gate moved. Link an optional field note to the research. On replay, change one condition and ask the player to make the next call. Start with five strategy players new to AI economics; target at least four who finish within 20 minutes and can explain, without prompting, why the lower list price did not settle the deal, when the reserved block becomes cheaper, and why the compact model initially failed. Record where they stopped or used a hint. This is a playtest target, not evidence of learning.

Before a playable build is called ready, run both successful paths and at least one failure in the actual browser, compare the ledger to this paper scenario, verify keyboard/touch input, and observe at least one playtest before expanding scope. The design puts the budget, two thresholds and one rival on the first screen; add systems when they change a decision the player can see.