Compute Power
The computational resources that support AI model training and inference, typically existing in the form of GPU clusters and data centers; it is a core indicator of an AI company's infrastructure capability and cost structure.
Core Facts
Computing power has become a barrier to entry for AI research, making it difficult for researchers at non-top-tier institutions to advance large-scale experimental validation due to limited resources.
In AI strategy formulation, the stability of compute supply and its cost should be treated as core variables.
Timeline (last 90 days)
From cloud computing to the mobile internet, and then to model training and inference, compute demand continues to expand over the long cycle.
The logic of computing power as an asset means that enterprises will reflect their computing power investment through capital expenditure (CapEx), depreciation cycles, and long-term purchase commitments.
Model parameters are no longer the only barrier, and sustainably obtainable computing resources are being repriced.
Compute has previously struggled to form a liquid market, and the core reason lies in its non-standardization.
At small batch sizes the bottleneck is memory bandwidth, while at large batch sizes the rising compute density makes compute the bottleneck, which explains the performance crossover between the 3090 and the L40S.
Large language models are a "flow-type" technology whose utility is tied to continuous input, including massive compute, energy, and constantly refreshed data.
Attacking the Bitcoin network requires controlling more than half of the hash rate, and the cost rises in step with network size, so security itself becomes a continuously invested "sunk cost."
Self-replicating models consume dispersed computing power and electricity. Although energy consumption at a single point is low, the total energy consumption after scaling up may not be better than that of a data center, and may even be harder to count and manage.
The large-scale adoption of machine learning in recent years is the result of the combination of increased computing power, the accumulation of massive data, and algorithmic improvements.
There is a positive feedback loop among compute, data, and algorithms: more data trains better models, and better models can extract more value from data, which in turn attracts more data and compute investment.
10 more timeline events
All Facts (20)
Computing power has become a barrier to entry for AI research, making it difficult for researchers at non-top-tier institutions to advance large-scale experimental validation due to limited resources.
75%VerifiedIn AI strategy formulation, the stability of compute supply and its cost should be treated as core variables.
65%UnverifiedFrom cloud computing to the mobile internet, and then to model training and inference, compute demand continues to expand over the long cycle.
50%UnverifiedThe logic of computing power as an asset means that enterprises will reflect their computing power investment through capital expenditure (CapEx), depreciation cycles, and long-term purchase commitments.
50%UnverifiedModel parameters are no longer the only barrier, and sustainably obtainable computing resources are being repriced.
50%UnverifiedCompute has previously struggled to form a liquid market, and the core reason lies in its non-standardization.
50%UnverifiedAt small batch sizes the bottleneck is memory bandwidth, while at large batch sizes the rising compute density makes compute the bottleneck, which explains the performance crossover between the 3090 and the L40S.
50%UnverifiedLarge language models are a "flow-type" technology whose utility is tied to continuous input, including massive compute, energy, and constantly refreshed data.
50%UnverifiedAttacking the Bitcoin network requires controlling more than half of the hash rate, and the cost rises in step with network size, so security itself becomes a continuously invested "sunk cost."
50%UnverifiedSelf-replicating models consume dispersed computing power and electricity. Although energy consumption at a single point is low, the total energy consumption after scaling up may not be better than that of a data center, and may even be harder to count and manage.
50%UnverifiedThe large-scale adoption of machine learning in recent years is the result of the combination of increased computing power, the accumulation of massive data, and algorithmic improvements.
50%UnverifiedThere is a positive feedback loop among compute, data, and algorithms: more data trains better models, and better models can extract more value from data, which in turn attracts more data and compute investment.
50%UnverifiedThe author proposes treating compute like fissile material: track it, and restrict access to the large-scale compute sufficient to elevate AI beyond known safe levels.
50%UnverifiedThe computing power dedicated to science approximately doubles every 2 years.
50%UnverifiedIndividual developers' computing power consumption is mainly reflected at two levels: token costs for calling cloud-based large model APIs and the GPU computing power required for local deployment of open-source models.
50%UnverifiedFor individual developers, computing power is a scarcer resource than model capability, and the stronger the model, the more expensive the computing power.
50%UnverifiedThe proliferation speed of AI capabilities far exceeds that of nuclear materials, and 'computing power' as a key input factor is harder to control than uranium.
50%UnverifiedContinuously rising demand for computing power is an irreversible trend.
50%UnverifiedCompute has become a quasi-financial strategic asset that can be actively stockpiled and flexibly monetized, with owners able to dynamically switch its use like managing an investment portfolio.
50%UnverifiedCompanies that control the underlying elements of 'computing power + energy' are gaining increasingly higher strategic value and capital premiums.
50%