Bright Simons says Ghana's AI strategy got its compute targets wrong by orders of magnitude

What Bright Simons found

Bright Simons, vice-president and head of research at the IMANI Centre for Policy and Education in Accra, has accused Ghana's Ministry of Communication, Digital Technology and Innovations of publishing compute targets in its national AI strategy that make no sense as written. In a long thread on X on 25 October 2026, reported the same day by MyJoyOnline, he said the strategy sets Ghana's total computing capacity target for 2027 at “1021 FLOPS” and its 2035 target at “1025 FLOPS”.

Read literally, he wrote, those figures would make the country's entire 2027 compute budget roughly equal to a 1970s desktop computer. His explanation is that someone meant ten to the power of 21 and ten to the power of 25 floating-point operations per second, and the superscripts were lost. His larger point is that nobody in the review chain caught it, through the ministry's own review, cabinet approval and a high-profile launch.

The thread follows a public debate between Simons and the minister, Samuel Nartey George, at JoyNews' Digital Economy Forum on AI and the Ghana Economy on Wednesday 21 October.

What the strategy actually says

AI Lookout checked the Republic of Ghana National Artificial Intelligence Strategy 2025-2035, a ministry document dated December 2025. On page 13, under Pillar 3 on digital infrastructure, recommendation 7 reads: increase Ghana's total compute infrastructure to 1021 FLOPS by 2027, increased by 0.5 OOM, meaning order of magnitude, every year, ensuring 1025 FLOPS total capacity by 2035. The same wording appears in the implementation table on page 50. In the rendered PDF the numbers are printed as plain four-digit figures, with no superscript.

Recommendation 9 on the same page calls for Ghana to drastically increase its energy capacity by 0.5 to 1 order of magnitude per year to support compute clusters. Page 68 says Ghana should produce at least one private AGI lab, defined as an AI lab with access to more than 1010 compute, with no unit given, and more than 10 billion cedis in mixed funding.

Elsewhere the strategy is ambitious in money terms too. It plans a National AI Fund starting at 5 billion cedis for 2025 to 2030 and growing to 15 billion, aims to attract 200 billion cedis of investment, and says AI should contribute 500 billion cedis to GDP by 2035.

Why the numbers do not add up either way

Simons argues the corrected reading is no better. Taking 10²¹ FLOPS at face value, he estimated that Ghana would need about 200,000 of NVIDIA's newest chips and 500 megawatts of power devoted to AI by 2027, at a cost of roughly $24 billion, which he compared to close to the government's entire programmed spending for 2026. The $250 million data centre now being procured, he said, would supply about 1% of that. For frontier model training he put the requirement higher still, at around 400,000 B300 GPUs and 1,114 MW, about a quarter of Ghana's peak demand.

He also tried the charitable reading that the drafters meant total training operations, not operations per second. On that view, 10²¹ operations is about 12 days of work for a single H100, and 10²⁵ is one GPT-4-sized training run, so the planned centre would hit the 2035 target in its first year. Either way, he concluded, the target is not meaningful.

The energy line draws the same objection. Growing power output by half an order of magnitude a year for eight years would be a ten-thousand-fold expansion, and Simons said he expects the energy ministry to deny making any such commitment. He called the episode “classic KATANOMICS”, his term for policies that proceed without independent critical review.

The $270 million GhanaChat plan

The real money is in GhanaChat. The strategy, under Pillar 8 on public-sector adoption, calls for a large language model trained on data from government agencies, used privately by government so that staff do not send confidential information to foreign AI systems, and recommends a mandate for all agencies to use it. In a separate MyJoyOnline report on the subsidy question, Simons said the $270 million plan, a data centre plus an AI fund, would buy about 2,100 GPUs and 4.3 MW of IT load, close to 90% of Ghana's entire commercial data-centre load in 2024.

By his modelling the centre would cost about $76 million a year to run, and a utilised GPU-hour would cost about $22 against a market rate of $6.71, implying a subsidy of roughly $61 million a year. These are his estimates, and MyJoyOnline notes that they would need independent technical and financial assessment.

The minister's case

At the forum, George outlined the $270 million plan: $250 million for infrastructure and compute and $20 million as seed money for an AI fund. His argument is about data. He said he had seen analysts in ministers' offices feed memos and sensitive government material into ChatGPT and Claude, calling it a critical national risk, and said Ghana's new data residency law is useless if government itself relies on a foreign model to process that data. That is the same worry behind Sweden's block on ChatGPT and Claude for government staff this week.

George admitted GhanaChat would not match ChatGPT or Claude at the start, but said it would improve with use. In remarks on cloud contracts, he called foreign providers' data-sovereignty assurances a mirage, citing the US CLOUD Act and his own exchanges with Google and Amazon Web Services. He described the questions raised as legitimate, while insisting the government is on the right track.

A cheaper alternative

Simons does not dispute the data problem. His proposal is to start with a sovereign AI assistant for the civil service rather than a national rival to the big chatbots. He estimated it would need about 144 GPUs, around 7×10¹⁷ operations a second at 8-bit precision, costing $13 million up front and $6.7 million a year. Instead of training a base model, he suggested hosting government data locally and switching between capable open-weight models through an orchestration layer.

Earlier in the week, Current Issues Online reported that he had also raised electricity and water at the forum, arguing that AI centres compete at roughly 3 to 5 US cents per kilowatt-hour while Ghana's power costs 10 to 12 cents, and asking what had changed since earlier local-software efforts failed.

What it means for national AI plans

Ghana's debate is a small-scale version of a choice many governments face. Building a national model promises control, as European leaders like Mistral's Arthur Mensch argue, but frontier-scale compute is expensive and power-hungry. Hosting open-weight models on local infrastructure can deliver most of the data residency benefit at a fraction of the cost. And as our report on countries hedging between US and Chinese AI blocs noted, Ghana is already among the top recipients of Chinese AI-related finance, which adds a geopolitical layer to any chip or model choice.

The FLOPS mistake is also a reminder that the unit matters. Compute capacity is measured in operations per second, while the size of a training run is a total count of operations. Confusing the two, or dropping an exponent, can shift a target by many orders of magnitude.

Limits and open questions

The strategy text is a primary source, and the misprinted figures are visible in it. The cost, power and GPU estimates, however, are Simons' own modelling and have not been independently verified. The ministry had not published a response to the FLOPS point when AI Lookout checked on 25 October 2026, so it is not known whether officials accept the error, plan a corrected edition, or intended a different unit.

It is also not clear how closely the $250 million centre now being procured follows the strategy's numeric targets, or whether the GhanaChat mandate will be implemented as written.

Why it matters

Ghana is spending real money on sovereign AI, and its founding document contains compute and energy targets that are wrong as printed and implausible when corrected. The underlying goal, keeping sensitive government data off foreign chatbots, is shared by critics and the minister alike. The open question is whether a $270 million national model is the right way to get there, or whether a smaller, locally hosted assistant would do the job first.