From the Real-World Operational Survival of Data Centers: The Cold Start Strategy of KAI.com

Let us dismantle and analyze this strategy starting directly from the real-world operational survival conditions of data centers.

THE COLD START STRATEGY FOR CONNECTING THE FIRST COMMERCIAL DATA CENTER TO KAI.COM

First, Understand the Real Pain Points of Data Centers

Realty、Iron Mountain、NTT、EdgeConnex、CyrusOne等）的典型现

Typical current status of global data center operators (Equinix, Digital Realty, Iron Mountain, NTT, EdgeConnex, CyrusOne, etc.): 痛苦 / Pain Point 程度 / Severity 根因 / Root Cause

GPU utilization is generally 40-60% 🔴 致命 / Fatal

Customers only purchase peak guarantees, leaving massive idle capacity during normal times.

Power contracts are long- term and non-refundable 🔴 致命 / Fatal

Signing a 20MW power contract but only using 12MW means paying base fees for 8MW in vain.

Direct sales teams cannot cover small-and-medium AI clients 🟡 中等 / Medium

Only able to serve tech giants (Microsoft, Meta, OpenAI), missing out on a massive long-tail market.

Inability to implement dynamic pricing 🟡 中等 / Medium

Contracts are fixed monthly/annually; unable to temporarily raise prices during sudden demand spikes.

Long queues during peak periods, completely empty during valleys 🟢 常见 / Common

Complete lack of any smoothing mechanism. KAI.com Strategy Document | 战略规划报告

Core Contradiction: The cost structure of data centers is fixed (power, cooling, land, O&M), but revenue is volatile. Every 10% increase in utilization boosts profit margins by 15-20%. This is a classic liquidity problem: “Sellers are willing to offer discounts, but cannot find buyers willing to purchase at that exact moment.”

Icebreaker Tool: Not “Joining an Exchange,” but “Activating Idle Assets”

Pitch to the First Data Center: Do not approach them by saying “Join our global computing power exchange.” It is too abstract and threatening (sounding like you intend to replace their internal sales team). Instead, you should say:

“Your H100 cluster utilization is currently at 55%. That remaining 45% of idle capacity is burning electricity bills every day. We have a way to turn this idle capacity into incremental revenue— without requiring you to modify any contracts, without you needing to contact new clients, and without changing your pricing strategy. What we do: We help you put this 45% idle capacity—note, strictly idle capacity— into a global liquidity pool. We will notify you when someone buys. When no one buys, you don’t have to do anything. The price is set by us, and it will never be lower than the minimum price you can accept. We do not touch your direct customers. We will never preempt computing power when your contract customers have demand. We only generate incremental value—our clients’ clients represent new demand actively explored by KAI.com.”

Key Design Principle: KAI.com builds complementary liquidity, not alternative liquidity. Data centers will never hand over their core capacity to you— that is their lifeblood. But they are highly willing to monetize marginal capacity (idle windows during nights, weekends, and quarter-ends). Just like airlines are willing to sell the last row of seats at a discount to Priceline without affecting full-price sales for front rows. KAI.com Strategy Document | 战略规划报告

Step 1: Selecting the First Data Center Requires a Strategic Approach

We cannot choose randomly. We must select the one that is most desperate for incremental revenue and experiences the highest pain. 筛选条件 / Screening Criteria 原因 / Reason

GPU cluster utilization < 50%

The operational pain point and financial pressure are the strongest.

Outside the sphere of influence of hyperscale cloud vendors

Less vulnerable to competitive squeezing or displacement by AWS/Azure.

Newly constructed clusters that are not yet filled with clients

Demonstrates the highest immediate requirement for cash inflows.

Geographical location possesses “timezone arbitrage” potential

Can effectively align with the cyclical and tidal demand of subsequent global clients.

Management/Founder has a VC or internet background

Exhibits a significantly higher level of openness and adaptability to innovations.

Most Likely First Candidate: Data centers located in the Nordics (Iceland/Norway/Sweden) or Quebec, Canada. Why them: Green power is exceptionally cheap (hydro/geothermal: $0.03-0.05/kWh). GPU operational costs are naturally 30-40% lower than Frankfurt/London. However, remote geographical locations lead to few local clients and low utilization rates. They are highly desperate to convert “cheap electricity” into “computing power exports.” This perfectly aligns with the national digital strategy directions of Nordic countries. Specific Target Companies: Norlandia / DigiPlex / Bahnhof / Green Mountain / Hydro66 (Small Nordic DC operators) or Canada’s Hydro Québec DC / eStruxture—which similarly offer cheap hydro but are geographically remote. • • • • • • • • • • KAI.com Strategy Document | 战略规划报告

Step 2: KAI.com Brings the First Customer to Transact

The golden rule of a cold start: Do not pull supply first and then look for demand. Pretend there is demand first, then pull supply. KAI.com cannot afford to wait for the data center to go online before looking for the first buyer—otherwise, the data center’s idle capacity will continue to burn cash during the waiting period. Correct Sequence: KAI.com directly rents spot instances on commercial hyperscale clouds (AWS/Azure/GCP). Train an actual proprietary model under KAI.com’s name (e.g., fine-tuning an open-source LLaMA). During training, KAI.com accumulates empirical scheduling data regarding which cloud is cheap and when. Leverage this data to approach the Nordic data center and pitch: “Look, I rented 1,000 H100 GPUs on AWS and spent $4.12/ hour during US daytime. If you sell your idle capacity from 8 PM to 8 AM to me for $1.80/hour, I will immediately migrate those 1,000 GPUs from AWS to you. Your additional revenue: $1.80 × 1,000 GPUs × 12 hours × 30 days = $648,000/month.”

Numerical Calculation: • 1,000 GPUs × 12 hours/day × 30 days = 360,000 GPU- hours/month • Market price $4.12/GPU-hour → absorbed by AWS • KAI offer $1.80/GPU-hour → Data center additional revenue $648,000/month • DC Cost (power + O&M) ≈ $0.40/GPU-hour → Gross profit $1.40/GPU-hour → $504,000/month pure extra profit For the data center: This is pure incremental revenue. These GPUs would have idled tonight anyway. For KAI.com: Securing computing power at $1.80 allows flipping it to the next customer for $2.50-$3.00.

The essence of a cold start is: KAI.com acts as the first market maker, using its own capital to carry inventory. Just like Citadel buys stock into inventory with its own money before making markets, and then sells when buyers arrive. 1. 2. 3. 4. 1. 2. 3. 4. KAI.com Strategy Document | 战略规划报告

Step 3: Designing a “Zero-Risk Integration” Technical and Commercial Solution

For a data center CTO, integrating a new scheduling system raises two primary fears: Security concerns—Will KAI.com breach or compromise their internal management plane? Contractual/Legal risk—If KAI.com’s clients default on payments, is the data center legally liable? KAI.com’s Solution Design:

Technical Layer Do not install any software agent inside the data center. Zero access required to the data center’s core management APIs. Integrate strictly via standard OpenStack / VMware / Kubernetes APIs. KAI.com’s scheduling layer only manages: “When, how many GPUs, and what image to run,” acting identically to a customer manually submitting a ticket in the DC Portal. Security Isolation: KAI.com rents isolated bare-metal tenant- level access.

Commercial Layer Pledge Model—KAI.com advances a one-month rent deposit. If KAI.com’s customers default, the amount is deducted from the deposit; the data center bears zero credit risk. Minimum Purchase Commitment—First month: KAI.com guarantees filling 500 GPUs for 24 hours. Regardless of downstream buyers, KAI.com carries the inventory risk itself. Price Floor Contract—Establish a floor price (the lowest price acceptable to the DC). When the market price is higher than the floor, the DC receives a higher revenue share; when below, KAI.com subsidizes the difference. The DC never loses.

This is the most classic “floor contract” in financial markets—you are selling a “no-loss” option to the data center in exchange for its liquidity. This option fee constitutes KAI.com’s upfront capital investment. 1. 2. 1. 2. • • • • • • • • • • • • • • • • KAI.com Strategy Document | 战略规划报告

Step 4: Using the First Data Center to Leverage the Second

Once the first data center is connected and KAI.com successfully brings it $500k+ in incremental monthly revenue, subsequent expansion enters a positive flywheel of network effects: Pitch to the Second Data Center: “We already have Nordic DC1 online with a monthly transaction volume of $2 million. If you connect, your clients will see more choices → your utilization will also increase. Moreover, we have real scheduling data—we know which time slots and which computing powers are worth how much. Your cost of integration is 80% lower than DC1 (because we have standardized the SDK/API/contracts). Integration takes only 3 days, and incremental revenue begins within 2 weeks.” When the 3rd, 4th, and 10th data centers join: Larger liquidity pool → higher matching success rate → more buyers arrive → more sellers want to join = A standard two- sided network effect. KAI.com Strategy Document | 战略规划报告

Step 5: From “Idle Computing Power Auction” to a “True Market Maker Exchange”

In the first six months, the data center integration model is discretionary excess capacity—the data center autonomously decides how many GPUs to put into the KAI.com pool tonight and at what price. But this is not enough. A true market maker requires continuous, guaranteed two-way quotes. When KAI.com scales to 10+ data centers and $20M+ monthly transaction volume, it launches a new model: “KAI.com Market Maker Custody” KAI.com leases a fixed number of GPUs long-term from data centers (e.g., 3,000 GPUs on a monthly subscription). KAI.com absorbs the idle risk but gains complete scheduling rights over these GPUs. It can act like a true financial market maker, posting bid-ask quotes at any time. Benefits to the Data Center: Revenue upgrades from “filling operational gaps” to a “stable, predictable monthly lease contract.” The price is slightly lower than selling to a single long-term lease customer, but there is zero risk of customer churn. The data center only collects revenue, delivers power, and handles O&M, without worrying about customer acquisition. Benefits to KAI.com: Possesses actual inventory, can quote real two-way prices, and acts as an active market maker rather than a passive matchmaking platform.

COMPLETE COLD START TIMELINE

Timeline

Core Strategy (EN)

T-1 Month

KAI.com rents spot instances on AWS/GCP to run real training; accumulates scheduling data + price spread analysis reports.

T+0 Month

Approach Nordic DC1 (Iceland/Norway). Present price spread analysis showing how much idle capacity is worth. Sign “floor contract” involving KAI upfront deposit + minimum purchase commitment. Discuss only “monetizing incremental idle capacity,” avoiding threats to direct sales. KAI.com Strategy Document | 战略规划报告

Timeline

Core Strategy (EN)

T+1 Month

DC1 goes online; the first batch of 1,000 H100 GPUs connects. KAI.com acts as the first buyer (running open-source model training) to validate technical pipelines, settlement workflows, and latency SLAs.

T+2 Month

KAI.com finds the first external buyer (a European AI company needing cheap computing power for multilingual model training). Match: DC1 night idle at $1.80/GPU-hour → sold to customer at $2.40. KAI.com earns a spread of $0.60/GPU-hour = $432,000/month gross profit.

T+4 Month

DC1’s monthly incremental revenue reaches $650k. KAI.com uses this case study to approach DC2 (Quebec, Canada). Pitch: “20% of DC1’s revenue growth comes from us. Do you want to try?”

T+6 Month

5 data centers integrated. Monthly transaction volume reaches $8 million. The exchange establishes basic functional liquidity.

T+12 Month

15 data centers integrated. Monthly transaction volume reaches $50 million. Launch “KAI.com Market Maker Custody” model. Begin carrying fixed inventory and offering continuous two-way quotes.

T+18 Month

Launch computing power futures contracts (30-day / 90-day forwards). The first customer locks in a 10,000-GPU training demand for Q2 next year. KAI.com uses futures to lock in long-term data center capacity, forming a true computing power derivatives market.

THE MOST CRITICAL EARLY METRICS

It is not transaction volume—it is the inventory turnover rate.

Inventory Turnover Rate = Monthly Sold GPU-hours / Average Custody GPU-hours KAI.com Strategy Document | 战略规划报告

• > 1.0 = Good (GPUs leased by KAI.com are successfully sold) • > 1.5 = Very Good • > 2.0 = Excellent (The market maker is utilizing capital highly efficiently) If turnover rate < 0.5 for three consecutive months: Indicates that the demand side is lagging. Need to boost marketing / lower prices / or reduce custody inventory scale.

Market Maker’s Profit = Spread × Turnover Rate × Inventory Scale. In financial market making, Citadel’s turnover can reach 50 times a day—but GPU market making cannot be that fast due to physical scheduling and deployment latency. A monthly turnover of 2-3 times is already top-tier. The corresponding annualized ROI = 15% spread × 24x turnover (2x monthly) = 360% annualized return on capital.

Summary in One Sentence: The first data center isn’t convinced by a grand “global computing power exchange”; it is convinced by “I will turn the 1,000 GPUs you have empty tonight into money.”

KAI.com first uses AWS/Azure spot instances to manufacture “synthetic inventory” and verify the existence of the spread; then uses its own capital to bear the idle risk of the first data center; eliminates data center fears with a floor contract; and gradually upgrades to a true market-maker custody model once real liquidity arrives.

Just like the history of any exchange—when the Chicago Board of Trade (CBOT) was founded in 1848, the first batch of farmer “sellers” also trusted a single person: “Put your corn with me, and I’ll find you a buyer.” CBOT itself didn’t grow a single kernel of corn or consume a single kernel of corn, but global corn pricing ended up in its hands. KAI.com’s computing power exchange is the exact same story—starting from the trust of the very first “idle GPU.” KAI.com Strategy Document | 战略规划报告