In the rapidly expanding artificial intelligence sector, large vertically integrated tech conglomerates own both the hyperscale physical GPU data centres (which lease out wholesale raw compute capacity) and downstream consumer-facing AI API services (which sell fine-tuned language model queries directly to developers). Independent AI application developers must lease wholesale GPU compute time from these integrated providers to run and deploy their own competing downstream models.
Recently, competition authorities have launched investigations into allegations that these dominant integrated providers are engaging in anti-competitive margin squeezes to eliminate downstream software rivals.
Fig. 2 – Wholesale GPU Compute – average revenue and average cost per 1,000 GPU-hours ($)
| Quarter | Average revenue per 1,000 GPU-hours ($) | Average cost per 1,000 GPU-hours ($) |
|---|---|---|
| Q1-21 | 120 | 100 |
| Q3-21 | 130 | 110 |
| Q1-22 | 130 | 115 |
| Q3-22 | 140 | 115 |
| Q1-23 | 220 | 95 |
| Q3-23 | 240 | 90 |
| Q1-24 | 235 | 85 |
| Q3-24 | 230 | 80 |
Fig. 3 – Downstream AI API Service – average revenue and average cost per million tokens ($)
| Quarter | Average revenue per million tokens ($) | Average cost per million tokens ($) |
|---|---|---|
| Q1-21 | 45 | 40 |
| Q3-21 | 48 | 42 |
| Q1-22 | 48 | 43 |
| Q3-22 | 50 | 44 |
| Q1-23 | 30 | 38 |
| Q3-23 | 28 | 39 |
| Q1-24 | 25 | 38 |
| Q3-24 | 24 | 38 |
Integrated cloud provider conglomerates were accused of actively executing strategies to exclude independent AI applications from the market in 2023 and 2024, compared to more cooperative market conditions in 2021 and 2022.
Using Figs 2 and 3, explain what evidence there is to support this view.
