10 Global Biopharma R&D Computing Power Case Studies — The Critical Key of Tidal Scheduling 行业特殊性 / Industry Specificities

The computational demand curve for biopharmaceutical R&D is the most extreme among all global industries: 特征 / Feature 表现 / Manifestation 对算力调度的含义 / Implications for Compute Scheduling

Pulsed Spikes

A single molecular docking simulation requires 5,000 GPUs × 72 hours non-stop.

Requires mobilizing massive compute in a short window -> Release immediately upon completion.

High Phase- Dependency

Target Discovery → Lead Compound → Preclinical → Phase I-III Clinical → Submission.

Compute demands vary by 10x to 100x between different R&D phases.

Complex Geo- Compliance

HIPAA (US), GDPR (Europe), Human Genetic Resources Regulation (China).

Data cannot cross borders freely, but compute scheduling can prioritize “local/nearby” resources. deadline刚性 Rigid Deadlines

FDA submission deadlines, ASCO annual meetings, patent expiration dates.

Compute demand climbs exponentially 3 months prior to the deadline.

Mixed Ecosystems

AlphaFold is open-source, but molecular dynamics simulations rely on Schrödinger’s proprietary software.

Different software stacks require distinct and optimized GPU configurations.

Core Pain Point: No pharmaceutical company will build its internal compute capacity to match peak demand. Peaks last only 3-6 months, whereas building a data center is a 10-year capital asset investment. This imbalance provides the fertile ground for KAI.com’s market-making. KAI.com | Global Biopharma R&D Computing Power Report 算力调度案例 / Compute Scheduling Case Studies

Company & Direction: Pfizer Inc. (New York, US), mRNA platform technology expansion (Influenza + RSV + Shingles + Cancer Vaccines). 算力需求画像 / COMPUTE DEMAND PROFILE

Normal Stage baseline load is 1,500 GPUs (500 for mRNA sequence design AI, 800 for LNP molecular simulation, 200 for stability prediction), covered by internal data centers + Azure. During a pandemic or outbreak stage, where new vaccines must be designed and enter clinical trials within 100 days, high-throughput screening and immunogenicity prediction cause total GPU demand to skyrocket to 8,000 cards—creating a 430% internal deficit. This “black swan pulse” peak lasts only 3-6 months. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

Under normal conditions, KAI.com routes remote backups and non-sensitive cross-border sequence comparisons to Iceland’s green power pool and Singapore’s nighttime idle compute, saving Pfizer $2.8M/month. Upon emergency vaccine launch, KAI.com mobilizes 6,500 GPUs within 3 hours: 3,000 from North American idle pools (midnight local time, $2.10/GPU-hr, 40% below Azure), 2,000 from European pools (afternoon relay of US midnight, GDPR compliant), and 1,500 from APAC pools (nighttime batch processing of anonymized data). 做市商价值 / MARKET MAKER VALUE

Pfizer’s API simply returns “8,000 GPUs allocated, avg $2.35/GPU-hr”, a 43% savings against Azure on-demand ($4.12), saving $7.6M for this emergency cycle. KAI.com delivers real-time liquidity across multiple geo-compliant zones rather than just cheap wholesale compute. KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: Novartis AG (Basel, Switzerland), Solid Tumor CAR-T, Radioligand Therapy. 算力需求画像 / COMPUTE DEMAND PROFILE

Data analysis demands for Phase I-III clinical trials are highly erratic. If CAR-T enrollment accelerates, Month 6 might suddenly require the analysis of 5,000 PET-CT images. AI image inference (segmentation, measurement, tracking) takes 5 minutes per image, totaling 25,000 GPU-minutes, requiring 35 GPUs for 12 continuous hours. The pain point is that localized clinical data across 42 countries cannot leave national borders, yet local sites only possess CPU workstations. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

KAI.com deploys a lightweight KAI Agent to detect local queues and executes “Local Inference, Global Scheduling”. When the Boston site submits 500 images, sensitive data is routed to a partner room at Mass General Hospital (40 idle cards, HIPAA compliant, stays in the US), while anonymized data goes to Montreal, Canada (90 cards, $1.50/GPU-hr, HIPAA compliant), finishing within 15 minutes. For Tokyo (PMDA compliant) submitting 3,000 images, it unifies Tokyo Equinix TY2 (120 cards), Osaka nighttime pools (80 cards), and anonymized Singapore backup (50 cards). 做市商价值 / MARKET MAKER VALUE

Novartis manages 72 IT vendors (AWS, Azure, local DCs). KAI.com acts as a unified liquidity layer across all vendors. The Novartis CTO doesn’t manage multiple environments, but simply requests: “Analyze 3,000 images, budget cap $5,000, compliance zones US and Japan, fastest turnaround.” KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: BeiGene (Beijing/Cambridge/Basel), BTK inhibitor iteration, bispecific antibodies. 算力需求画像 / COMPUTE DEMAND PROFILE

R&D centers span Beijing, Shanghai, Suzhou, Guangzhou, Cambridge (US), and Basel (Switzerland), submitting data simultaneously to NMPA, FDA, and EMA. Shanghai Zhangjiang runs molecular dynamics, Cambridge runs AI antibody design, and Guangzhou runs process optimization. Under China’s Human Genetic Resources Administration regulations, raw biological data cannot cross borders, yet algorithm models require shared training—a classic Federated Learning scenario. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

KAI.com established a “Three-Zone Federated Computing Layer”. China Zone (zero data exfiltration) routes daytime molecular simulations to Shanghai (close to algorithm teams) and shifts nighttime model training to Zhangjiakou/Ulanqab (cheap green power). US Zone (Cambridge/Princeton) uses local Boston compute by day and Virginia/Ohio DCs by night, transmitting only encrypted model parameters back to China. Europe Zone (Basel) runs clinical analysis locally, scaling to German/French idle capacity at night under GDPR frameworks. 做市商价值 / MARKET MAKER VALUE

KAI.com performs “model parameter market-making” across zones. If China’s nighttime compute ($1.20) is cheaper than the US ($2.40), the federated aggregation task routes to Beijing; raw data never moves, only encrypted gradients transfer. This solves geopolitical compute compliance blocks. KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: Moderna (Cambridge, MA, US), mRNA-4157 personalized cancer vaccine (in partnership with Merck). 算力需求画像 / COMPUTE DEMAND PROFILE

Personalized cancer vaccines require an individualized cycle per patient: tumor biopsy → whole-exome sequencing → neoantigen identification → mRNA sequence design, which must be finalized within 4-6 weeks. A single patient requires 50 GPUs running for 72 hours (3,600 GPU-hours). If a clinical trial enrolls 1,000 patients concurrently, it averages 85,714 GPU-hours per day—requiring 3,571 GPUs running 24/7. Moderna’s internal capacity is ~2,000 H100 GPUs, leaving a 44% spike deficit. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

When clinicians submit 200 samples at once and overload internal capacity, KAI.com flags the remaining 88 samples as top-priority “premium liquidity demand.” It instantly extracts capacity from connected networks: 1,000 GPUs from a reserved “medical compute spare pool” (contractual 48-hour reclamation), 500 GPUs from North American midnight pools, 300 GPUs from European post-peak afternoon pools, and sweeps 400 GPUs from AWS spot markets by bidding $0.10 higher, assembling 2,200 extra GPUs. 做市商价值 / MARKET MAKER VALUE

KAI.com quotes $2.80/GPU-hr during peaks (32% below AWS on-demand) and drops to $1.80/GPU-hr during troughs, smoothly guiding non-urgent research to off-peak periods. Unlike unpredictable cloud spot instances, KAI.com’s “market maker inventory” provides absolute contractual guarantees, empowering Moderna to race against time. KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: Takeda Pharmaceuticals (Osaka, Japan), rare genetic diseases, AI-driven drug repurposing. 算力需求画像 / COMPUTE DEMAND PROFILE

Of 8,000 known rare diseases, only 5% have treatments. Traditional paths cost $2.6B and take 10-15 years, while AI drug repurposing shortens this to 2-3 years. Takeda’s AI engine cross-references approved drug molecular structures, disease targets, and real-world patient records to run molecular docking simulations for every drug-disease combination—totaling 160 billion simulations, demanding 50,000 GPU-years. Takeda’s global centers own only 8,000 A100 GPUs, leaving an 85% gap. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

Restricted by Japanese health data rules, European GDPR, and US HIPAA, KAI.com deploys a distributed federated model. Japan Zone: routes Tokyo Equinix TY2 nighttime idle capacity (~$1.30/GPU-hr) and Hokkaido’s green-powered cool data centers ($1.10/ GPU-hr) to back up Osaka, eliminating data export risks. Europe Zone: links Vienna and Cambridge centers with central idle pools in Frankfurt/Zurich, boosting utilization by 30%. US Zone: draws from North American nighttime troughs. Model parameters are aggregated across regions via KAI.com’s encrypted transport layer. 做市商价值 / MARKET MAKER VALUE

The same project advances simultaneously across three zones, with the system automatically electing the cheapest available zone as the aggregation node. Takeda tracks progress and a global weighted average price of $1.90/GPU-hr on a single dashboard, saving 38% compared to single-region sourcing while seamlessly aligning geo-compliance with optimal cost efficiency. KAI.com | Global Biopharma R&D Computing Power Report

(Germany)

Company & Direction: BioNTech SE (Mainz, Germany), personalized neoantigen vaccines, African infectious diseases (Malaria, TB, HIV) mRNA vaccine development. 算力需求画像 / COMPUTE DEMAND PROFILE

BioNTech established mRNA manufacturing facilities in Rwanda, but localized computing infrastructure in Africa is practically non- existent. Its German hubs house ~3,000 GPUs dedicated to GDPR-bound European vaccine designs. With the African project active, they must analyze terabytes of genome sequences of local malaria parasite variants and optimize mRNA sequences. Due to high international latencies and African nations’ rigid mandates that “data must be analyzed within borders,” data back-transfer to Europe is unfeasible. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

KAI.com implements a “Proximity Merging of Compute and Data” approach: sequencing data from Rwanda is routed over a low- latency (15ms) fiber connection to Cape Town, South Africa (NTT/Teraco DCs house ~2,000 H100 GPUs, with a 35% idle rate yielding 700 available cards), utilizing lower African power and labor costs at $1.60/GPU-hr. Backup routes include Tel Aviv, Israel (40ms latency) and the Sines data center in Portugal via West African subsea cables, guaranteeing resilience under extreme scenarios. 做市商价值 / MARKET MAKER VALUE

Through KAI.com’s market-making capabilities, BioNTech acquires local African compute at $1.60/GPU-hr, saving 50% compared to the total cost of routing back to Europe ($3.20 compute fee + expensive cross-border bandwidth), while fully satisfying Rwandan local data policies. These price signals further attract computing infrastructure investments into emerging markets, advancing global health equity. KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: AstraZeneca (Cambridge/Gothenburg/Gaithersburg/Shanghai), next-generation and dual-payload Antibody- Drug Conjugates (ADCs). 算力需求画像 / COMPUTE DEMAND PROFILE

ADCs combine an antibody, a linker, and a payload. AI models must execute high-throughput virtual screening across a massive combinatorial space of 10^12 variations. This workload is highly parallelized and latency-insensitive, but cannot tolerate mid-run interruptions (which invalidate previous steps). It generates terabytes of data and experiences extreme spikes before major events like ASCO or ESMO. AstraZeneca distributes 2,000 (UK), 1,000 (Sweden), 1,500 (US), and 800 (China) GPUs globally under conflicting tariffs and rules. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

During the 3-month sprint leading up to ESMO, 5 million simulations require 50 million GPU-hours. KAI.com orchestrates a “global compute relay race”: Cambridge (UK) handles real-time tasks by day, while its nights/weekends are swept alongside 500 extra H100s from London Equinix LD4 into training. Gothenburg (Sweden) capitalizes on cheap Nordic hydropower ($0.04/kWh), with KAI.com scheduling 80% of massive training blocks into its low-tariff nights. Gaithersburg (US) acts as the daytime relay anchor during European nights. Shanghai (China) is restricted by human genetic rules, so KAI.com automatically filters and routes only non-Chinese origin data simulations there. 做市商价值 / MARKET MAKER VALUE

AstraZeneca’s infrastructure team previously grappled with 4 separate jurisdictions, contracts, and compliance rules. KAI.com abstracts this into a unified API. The system automatically fractures and dispatches workloads to nodes with the lowest real-time pricing globally, achieving a weighted mean of $2.10/GPU-hr—32% lower than standard single-region cloud pricing. KAI.com | Global Biopharma R&D Computing Power Report

(Switzerland/US)

Company & Direction: Roche/Genentech (Basel, Switzerland / South San Francisco, US), oncology, Real-World Evidence (RWE) driven drug development. 算力需求画像 / COMPUTE DEMAND PROFILE

Roche leverages RWE to extract clinical insights from 45 million patient electronic health records globally. This requires large-scale batch inference across petabytes of multimodal data (text records, lab metrics, radiology reports), with a complete model refresh every three months. RWE analysis has an intensely seasonal cycle: demand sextuples three months prior to FDA/EMA submission deadlines and two months before major oncology congresses, whereas baseline periods require only 30-40% of peak capacity. Roche maintains around 6,000 GPUs, with a baseline utilization rate below 50%. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

KAI.com implemented an “RWE Peak Hedging Model”: during low-utilization troughs (40%), Roche’s surplus capacity is leased through KAI.com to other pharma players (e.g., matching Bayer’s genetic tasks into Roche’s idle clusters). Roche earns a baseline monetization income of $0.80/GPU-hr, and Bayer acquires compute 30% below market rates. When Roche hits its sextupled peak for submission deadlines, KAI.com triggers its hedging playbook: Phase 1 triggers the contractually guaranteed reclamation of 2,500 shared GPUs from other users; Phase 2 spins up cheap hydropower compute from the US Pacific Northwest (Oregon/ Washington); Phase 3 activates the reserved medical spare compute pools in Europe. 做市商价值 / MARKET MAKER VALUE

This represents a true market maker form: the exact same batch of GPUs serves different enterprises across temporal cycles— Bayer utilizes them during troughs, and Roche reclaims them during peaks. KAI.com does not merely resell compute; it operates as a “computational central bank” that manages liquidity across multiple pharma giants, smoothing the aggregate demand curve. KAI.com | Global Biopharma R&D Computing Power Report

Korea)

Company & Direction: Samsung Biologics (Incheon, South Korea), CDMO (Contract Development and Manufacturing Organization), biosimilars, digital twin commercial manufacturing. 算力需求画像 / COMPUTE DEMAND PROFILE

Samsung Biologics manages over 300,000 liters of bioreactor assets. Its AI digital twin models collect 1,000+ data points per second (temperature, pH, dissolved oxygen, glucose levels) from every 10,000-liter tank to predict optimal yield windows and contamination risks, ingesting terabytes of production data daily. The computational profile exhibits a textbook production model: real-time process control requires 200 GPUs/day for low-latency inference, while end-to-end digital twin model refreshes require a training workload of 2,000 GPUs/month—a 10:1 training-to-inference ratio. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

South Korea provides world-class network bandwidth, but core server real estate in Seoul and Bundang is severely constrained, making expansion difficult alongside high power tariffs ($0.10-0.12/kWh). KAI.com executes a “Production Compute Tidal Stratification”: the non-interruptible 200-GPU real-time inference runs at the Incheon data center adjacent to the reactors. Samsung pays a reservation fee; when local load is low, KAI.com runs low-priority tasks, but reclaims all cards unconditionally if inference spikes. For offline training tasks that tolerate a 2-hour latency, KAI.com routes them to the lowest-cost windows: switching to Seoul’s off-peak nighttime power (40% savings), bridging via the Busan-Tsushima subsea cable to cheap facilities in Kyushu, Japan (15ms latency), or dispatching non-proprietary media optimization training to ultra-low-cost Southeast Asian nodes in Malaysia or Indonesia. 做市商价值 / MARKET MAKER VALUE

Samsung Biologics’ compute procurement shifts from a “fixed capital expenditure” model to an “elastic liquidity sourcing” model. KAI.com does not touch core operational control, but acts as an underlying “compute cost optimization layer,” utilizing elasticity and optimized cost structures to insulate production peaks. KAI.com | Global Biopharma R&D Computing Power Report

Company & Direction: Regeneron Pharmaceuticals (Tarrytown, NY, US), gene editing therapies, CRISPR guide RNA (gRNA) off- target prediction, whole exome association studies (GWAS). 算力需求画像 / COMPUTE DEMAND PROFILE

Regeneron operates a premier genetics center (RGC), having sequenced whole exomes for over 2 million individuals (40 trillion data points), requiring 2,000 GPUs for 30 consecutive days as baseline research. When a new CRISPR target launches, designing 100,000 candidate guide RNAs (gRNAs) and running 300 off-target predictions demands 1,000 GPUs for 5 intense days. Though Regeneron owns a top-3 pharma supercomputer (~5,000 GPUs), gene-editing runs are often triggered suddenly by scientific breakthroughs, requiring immediate, massive scale-up. KAI.COM潮汐调度方案 / KAI.COM TIDAL SCHEDULING SOLUTION

When the Chief Scientific Officer abruptly decides to initiate an editing simulation for a gene (e.g., FBN1), 10,000 GPUs are needed by the next morning, while their internal 5,000 GPUs are fully utilized by ongoing projects. KAI.com’s market-making engine delivers a flash response within 3 hours: Step 1 instantly releases 1,500 GPUs from KAI’s long-term inventory reserved for Regeneron; Step 2 sweeps midnight idle capacity across North American hubs; Step 3 exercises emergency drawdown rights from “compute futures contracts” purchased by Regeneron, pulling 3,500 GPUs from global low-cost nodes to secure 10,000 GPUs within 3 hours. 做市商价值 / MARKET MAKER VALUE

Regeneron’s average cost is controlled at $2.60/GPU-hr, 28% below standard on-demand cloud tariffs. The pharma giant is invoking a highly liquid, multi-layered compute spot and futures market. Layered over their internal supercomputer, KAI.com serves as a contractually guaranteed second liquidity tier that cannot be preempted or cut off mid-run. KAI.com | Global Biopharma R&D Computing Power Report 跨行业横切总结 / Cross-Industry Executive Summary

Company

Nation 核心场景 / Core Scenario KAI.com做市价值 / Market Maker Value

Elasticity Pfizer

Emergency mRNA R&D

Cross-continent 40x elasticity, automated geo- compliant backfills. 430% Novartis

CH

Clinical Image AI

Local distributed inference, unified layer across 72 IT vendors. 300% BeiGene

CN

Global Antibody R&D

Three-zone federated learning, encrypted model parameter routing. 200% Moderna

Personalized Vaccine

Market-maker inventory guarantees, peak/trough price hedging. 250% Takeda

Rare Disease AI

Global federated computing, multi-zone weighted price discovery. 600% BioNTech

Africa mRNA Initiative

Proximity bridging (South Africa), geo-compliance

• 50% cost cut. 500% AstraZeneca

UK-SE

ADC Molecular Design

24/7 global time-zone relay race, automated API billing. 400% Roche

CH

Real-World Evidence

Shared compute bank among pharma giants, peak/trough hedging. 600%

Bioreactor Twin

Layered production scheduling, hedging local tariffs with overseas nodes. 150% Regeneron

CRISPR gRNA Design

Flash spot aggregation paired with guaranteed “compute futures”. 200% KAI.com | Global Biopharma R&D Computing Power Report

Traditional hyperscale cloud providers are hardwired to orchestrate resources exclusively within their own “walled gardens.” AWS will never route workloads to Azure data centers, nor will GCP tap into a third-party Equinix facility. However, biopharma computational needs are inherently cross-cloud, cross-border, and cross-compliance. KAI.com builds the missing, sovereign-neutral liquidity layer across all walls. 2. 为什么药企自己不做？ / Why Can’t Pharma Giants Do It Interally?

While the top 20 pharmaceutical companies command an aggregate annual IT budget of ~$50B, these resources are fragmented across localized silos, countless POCs, and customized frameworks. No single enterprise can economically justify maintaining a multi-thousand-person global compute orchestration engineering team. KAI.com centralizes this market-making role, utilizing automated scheduling AI to manage 99% of tasks, driving efficiency thousands of times higher than traditional power grid paradigms. 核心总结 / One-Sentence Summary

“Genomes don’t wait, clinical trials don’t wait, and FDA deadlines don’t wait. Yet, massive computing clusters sit idle around the globe, waiting to be mobilized. KAI.com’s role as a market maker is to establish a non-stop, hyper-liquid marketplace that bridges the rigid timetables of life sciences with the transient, unutilized capacity of global infrastructure.” KAI.com | Global Biopharma R&D Computing Power Report