The Social History of Compute — Credit, Crypto, AI, and Space

CPUs, GPUs, memory, storage, networks, electricity, cooling, and operating time are finite resources that several uses claim at once. An “idle cycle” still carries a marginal cost in power and heat, and the owner is choosing between science, gaming, cryptocurrency, AI, business work, and doing nothing. Capacity does not flow to one purpose on its own. Software support, reward, story, cost, and institutional design allocate it.

A straight line in which benevolent science was taken over by Bitcoin, passed through Ethereum, and finally became AI erases the evidence. Bitcoin’s GPU era was a stage on the way to ASICs, AI was using GPUs before Ethereum existed, and Folding@home expanded dramatically under an urgent social goal long after cryptocurrency markets had appeared. Overlapping uses and different classes of hardware have to be kept apart.

This article separates documented institutions and dated observations from limited inferences and causal claims that remain unknown. “Human desire never stops” is not a historical fact that a source can verify. Evidence can show which systems measured, recognized, or paid for which behavior.

BOINC credit is an account of validated computational contribution. It is not money, ownership, voting power, or the scientific result itself. Total credit records what has accumulated, recent average credit (RAC) shows recent activity, and leaderboards for hosts, volunteers, and teams, along with third-party cross-project statistics, made the competition visible worldwide.

BOINC’s designers recorded that leaderboards pushed volunteers to add more PCs, upgrade hardware, or buy dedicated machines, and that team competition brought in friends, relatives, and coworkers. SETI@home’s retrospective likewise describes a culture built from a screensaver, a scientific purpose, a rank, and a team. Points were not decoration. They became a social institution for staying with the work and belonging to something.

Not every project went entirely unrewarded. GIMPS attributes discoveries and gives awards, and distributed.net shared prize money from the RC5 challenges. The accurate generalization is that routine work usually came back as scientific participation, credit, rank, badges, and certificates rather than as a tradable payment priced per job.

How credit is designed cannot be separated from integrity. Once a number matters, it invites inflated claims, modified clients, unauthorized machines, and selective returns. Validators, replication, passkeys, return-rate conditions, and fair-play rules grew up alongside the game. Replication and quorum remain project-specific rather than one universal method of validation.

Bitcoin did not put a price on scientific computation in general. It built a SHA-256d target search into block proposal and chain selection, and it assigns the block subsidy and transaction fees to a successful block. In a pool, a share at lower difficulty measures internal contribution; it is not itself a Bitcoin consensus block.

BOINC credit mostly records who contributed how much validated research computation. A Bitcoin reward records who satisfied a protocol condition and obtained a claim that can be transferred economically. A pool may look like a project server from the outside, but the purpose, the verifier, the reuse of the output, and the transferability of the reward all differ.

That difference can change an owner’s opportunity cost: the same electricity and the same machine might go to credit and scientific participation, to a reward the market prices, or to neither. But the fact that the choice exists as an institution is not a measurement of how many machines actually moved.

We found no cross-platform dataset that establishes such a causal claim. Active participation in SETI@home had already declined from its early publicity peak before Bitcoin appeared. A self-selected BOINC survey in 2006 recorded several reasons for leaving, among them computer problems, complexity, forgetting to restart, losing interest, and the cost of electricity, but it is not a representative sample and no causal shares can be calculated from it.

A 2012 account estimated about 900,000 active volunteer computers, roughly 10 PFLOPS, and GPUs supplying about 70 percent of the capacity. BOINC’s 2019 platform paper reports a snapshot from around 2018 of about 700,000 devices, 560,000 GPUs, and an average of 93 PFLOPS. The scope and the definitions differ, so those two dots do not form one trend line, and faster hardware separates the number of participants from FLOPS.

During the 2020 COVID-19 response, Folding@home reported that more than 700,000 citizen scientists joined over three weeks and that participation rose twentyfold; its project timeline records reaching roughly 1.5 exaFLOPS on 25 March 2020. That is neither sustained output nor directly comparable with BOINC, but it does counter the claim that cryptocurrency markets had made large-scale voluntary mobilization impossible.

CUDA arrived in 2006 and OpenCL 1.0 in 2008, widening software support for GPU work beyond graphics. BOINC broadened its GPU support in 2008, and SETI@home and GPUGrid distributed scientific applications. AlexNet was trained on two GeForce GTX 580 cards in 2012, so GPU-based AI research predates Ethereum.

Bitcoin saw a public OpenCL miner in 2010 and an FPGA implementation in 2011, and the first Avalon ASICs shipped in January 2013. Repetitive fixed SHA-256d work tied directly to revenue gave miners a strong reason to leave general-purpose GPUs for dedicated circuitry. Competitive Bitcoin mining today is led by SHA-256 ASICs; Bitcoin did not stay a consumer-GPU workload for long.

Science, gaming, Bitcoin, Ethereum, and AI overlapped rather than replacing one another in turn. Public evidence does not track how many GPUs left BOINC or Folding@home for cryptocurrency, so the width of a lane or the shape of a timeline cannot be read as market share.

Ethereum Frontier launched on 30 July 2015, and Mainnet produced blocks under Ethash proof of work. Ethash used a large dataset and a memory-hard design that kept GPU mining viable for years, but it did not shut ASICs out forever, and dedicated hardware appeared later. ASIC-resistant and ASIC-proof are different claims.

Company filings establish that cryptocurrency demand affected GPU markets. AMD’s 2017 Form 10-K recorded demand for GPUs suited to mining currencies including Ethereum. In Q3 2018 AMD said blockchain-related GPU sales had fallen from a high-single-digit percentage of total revenue in the prior-year quarter to negligible. NVIDIA’s FY2018 Form 10-K likewise recorded a large increase in mainstream OEM GeForce products aimed at cryptocurrency mining.

The same vendors disclosed that they had limited visibility into end use and could not measure the impact precisely. Demand, channel inventory, and resale effects can be observed. Assigning all of them to Ethereum alone, or to devices leaving BOINC, cannot.

Ethereum Mainnet switched to proof of stake in the Merge on 15 September 2022, which ended mining as a way of producing valid Mainnet blocks. ethereum.org estimates a reduction in energy consumption of about 99.95 percent. That was a consensus change for Ethereum Mainnet, not the end of other proof-of-work chains or of Bitcoin mining.

NVIDIA later said the Merge may have reduced the usefulness of GPUs for mining and increased resale into secondary markets, which could affect demand for lower-end products. The same disclosure listed macroeconomic conditions, China, and channel inventory among the other factors, so it does not support blaming a decline in gaming on the Merge alone.

A GPU that had been mining Ethereum could be sold, returned to gaming, pointed at another proof-of-work chain, put to work on compatible science or small AI jobs, or simply switched off. We found no primary dataset that measures where they went. “Mining GPUs became AI” turns a possibility about compatibility into a migration claim that nothing supports.

GPU use in AI did not begin abruptly after cryptocurrency. AlexNet reported training its 2012 ImageNet model on two GTX 580 GPUs for five to six days. Low-precision matrix operations, HBM, ECC, fast interconnects, collective communication, compilers, and frameworks later grew that approach into data-center systems for training and inference.

NVIDIA’s FY2026 Form 10-K reports Data Center revenue of $193.7 billion, total revenue of $215.9 billion, and Gaming revenue of $16.0 billion. Data Center includes networking and other products, so it is not all generative-AI GPU revenue. AMD’s FY2025 Form 10-K/A reports $16.6 billion of Data Center revenue and demand for EPYC and Instinct products, while Intel’s FY2025 Form 10-K says it missed the major shift in demand toward AI-optimized GPUs and had not become a meaningful participant through Gaudi.

“Now AI” is therefore defensible as a market emphasis visible in several companies’ product roadmaps and revenue mix. Outcomes differ sharply from company to company, and these filings cannot be added up into an industry-wide AI total. Nor does any of it mean that household mining boards physically became H100-class clusters. Consumer boards and data-center systems differ in memory capacity, precision, interconnect, reliability, software support, cooling, and operational guarantees.

FLOPS, hashes per second, tokens per second, and BOINC credit cannot be converted into one another. Even within a single silicon family, the application, precision, memory traffic, communication, and validation all differ. Allocating resources means asking which workload runs under which software and system boundary at which energy cost, not just counting devices.

Computing in space is not only a matter for the future. NASA says the first Spaceborne Computer, sent to the ISS in 2017, ran a roughly one-TFLOP COTS system for 207 days without a reset. For Spaceborne Computer-2, sent to the ISS in 2021, NASA’s 30 December 2022 operations report records the units being prepared for return and describes an experiment that tested onboard data processing, AI, and error mitigation.

ESA’s PhiSat-1 launched on 3 September 2020 and used an Intel Movidius Myriad 2 to classify clouds in Earth-observation images, discarding low-value images before downlink. That demonstrated edge AI close to the sensor. It was not a generally available orbital cloud data center.

NASA’s High-Performance Spaceflight Computing project is developing a new processor that combines higher performance with radiation tolerance. Its 2026 status update reports that chips have been manufactured and that testing is under way. It should not be described as space-qualified or deployed until those tests establish that status.

ESA’s Cognitive Cloud Computing in Space campaign separates application demonstrations on existing orbital hardware from short studies pointing toward a future networked-data-center vision. ESA’s 2023 review names prototypes and higher technology readiness as the work that comes next. These are real development steps, not evidence that a permanent public utility already stands in for terrestrial cloud infrastructure.

Space has no atmospheric convection, so heat moves through structures and leaves through radiators. Thermal design, single-event radiation effects, spacecraft power and mass budgets, and communication delay or disruption all sit inside the system boundary. Sunlight does not make electricity or cooling free.

No primary evidence found through 23 August 2026 showed a commercially operating Bitcoin-mining farm or BOINC data center in space. That does not prove such a system can never exist, but it cannot be presented as a current deployment either. One near-term use case that has been demonstrated is processing sensor data where it is produced, to reduce downlink and latency.

People run computers to take part in science, to be recognized, to belong somewhere, to compete, for income, out of curiosity, and for public purpose. Evidence can document behavior and institutions, but it cannot reduce human desire to one metric. A world-class library should keep the relationship traceable: who computed what, on which machine, at whose cost, under whose validation, and for whose benefit.