Slide contents
1. AI Leaders: Business Lab for Executives — HSE Faculty of Computer Science
AI Leaders: Business Lab for Executives — HSE Faculty of Computer Science
2. Evolution of AI Project Approaches
From infrastructure to GenAI applications on top of foundation models
Context — AI moved from server rooms to subscriptions.
Goal — Choose deployment, team, and launch economics.
Key Questions — Resources, scaling, TCO, and technical loop.
3. On-Premise Hardware
Past AI infrastructure
Large CAPEX for GPU servers.
Server room: power, cooling, security.
Long deployment cycles and complex scaling.
AI projects mostly fit large corporations.
4. Cloud Compute
Capacity without hardware purchase
AWS/Azure/GCP made resources subscription-based.
Pay-as-you-go: start without CAPEX.
Scale in minutes.
Constant high load can favor ownership.
5. ML Tooling Evolves
From scratch code to libraries
Open source: scikit-learn, TensorFlow, PyTorch.
Experiments and model development accelerate.
Pre-trained models lower entry barriers.
Specialists and ML pipelines remain necessary.
6. Foundation Models
API instead of training
GPT-3/ChatGPT cover many tasks.
Access via API, no training.
Text, speech, image as services.
Focus: tuning, prompting, integration.
7. GenAI Applications
Prompting → RAG → Agents
2023: prompting adapts the model.
2024: RAG adds knowledge and context.
2025: agents call tools.
2026: multi-agent handles complex work.
8. Roles Shrink to Tiny Teams
Specialists become hybrid teams
Before
Data engineer and ML researcher
ML engineer and DevOps
Software engineer for integration
Today
2–3 people for a pilot
Prompt engineer / AI orchestrator
MLOps + PM with AI background
9. Cost Estimation and Hidden Line Items
What matters when planning the budget
Infrastructure — On-prem — upfront costs; cloud — usage-based OPEX.
API and Scaling — Per-request cost grows with load.
Data and People — Data, labeling, expertise, and support drive cost.
10. Technical Side
Practical launch principles
Start with a clear-impact pilot.
Cloud default; on-prem for strict requirements.
Use ready-made services.
Design UX and errors around the model.
11. Support Bot TCO
200K dialogs/month, SLA under 5 sec
Input Parameters — 200K dialogs/month, 10 messages, 5 sec SLA.
Calculation Goal — Compare CAPEX/OPEX, team, and time-to-prod.
Solution Format — TCO view: load, model cost, sizing, final choice.
12. From Parameters to Load
10 messages = 5 bot responses
LLM calls: 1,000,000 responses/month.
Average load: ~0.386 RPS.
Peak hour: ~1.85 RPS.
5 sec SLA: ~9-10 parallel generations.
13. OpenAI gpt-5.2: Cost
Support response profile
Prices: input $1.75/M, output $14/M.
Profile: ~1,400 input / ~180 output tokens.
Month: 1.4B input, 180M output.
LLM total: ~$4,970/mo, ~$0.0249/dialog.
14. OpenAI API: CAPEX, OPEX, team
API advantage — fast start with minimal CAPEX
Financials
CAPEX: virtually 0.
LLM OPEX: ~$5k/mo (base profile).
Platform around it: typically $0.5k–$5k/mo.
Implementation
Launch: 5–7 FTE.
Operations: 1.5–3 FTE.
MVP: 4–6 weeks; production: 8–12 weeks.
15. Qwen Self-Host: 5 sec SLA
Peak load
Qwen-14B: ~86 tok/s; 32B: ~50.
180 tokens: ~2.1 / ~3.6 sec.
Parallelism: ~2 / ~1.
Peak ~10 responses: 14B 6-8; 32B 12.
16. Self-Host: CAPEX, OPEX, team
Tokens become infra + ops
Financials
14B: ~$120k–$220k; 32B: ~$200k–$350k+.
Amortized $180k: ~$5k/mo.
OPEX: colocation, maintenance, GPU ops.
Implementation
Launch: 8–12 FTE.
Operations: 3–5 FTE.
SLA: 12–20+ weeks plus delivery.
17. Qwen on Rented GPUs
Lower CAPEX, higher OPEX
Financials
CAPEX: nearly zero.
OPEX: hourly GPU rental + network + storage.
At high utilization, own hardware wins.
Implementation
Launch faster than buying servers.
Needs MLOps/SRE and inference control.
Risks: GPU price and availability.
18. GPU Rental OPEX
A100: 4.5 RUB/min
A100: 270 RUB/hour.
1 GPU 24×7: 194,400 RUB/mo.
14B (6–8 GPUs): 1.17–1.56M RUB/mo.
32B (~12 GPUs): 2.33M + network/reserve.
19. Three Deployment Strategies
200,000 dialogs per month
OpenAI API: fast start.
Own hardware: control and scale economics.
Rented GPUs: no CAPEX.
At current load, API-first wins.
20. Case recommendation
Without strict on-prem needs, start with API.
Rented GPUs: model control without large CAPEX.
Own hardware wins at high stable traffic.
API economics: output length and cacheable input.
Recommendation: launch via OpenAI API and review unit economics regularly.
21. Materials
API, GPU, self-host
OpenAI API docs/pricing
cloud.ru GPU pricing
Hugging Face model/deploy docs
FinOps unit economics
22. Key Conclusions
AI is commodity: key capabilities are cloud services.
Deploy in cloud; hire small core + domain.
Calculate economics upfront: scaling, API, data, people.
Success combines technology and competent management.
Use proven lessons for durable business value.