Dell survey shows data storage, not GPU, is stalling AI in production in Brazil

- Reading
- 5 min
- Tables
- 1
A study commissioned by Dell with 100 Brazilian executives shows that the brake on AI in production isn't a lack of chips: it's slow data, fragmented infrastructure, and an energy bill the board can no longer ignore.
The pilot works, production stalls
The script repeats itself across much of corporate Brazil: the generative AI pilot impresses in testing, leadership approves the investment, and the technology department is ordered to roll it out across the entire operation. It's in that leap, from lab to production, that the numbers stop adding up. According to the Modern Enterprise Readiness Study, a study commissioned by Dell Technologies and conducted by the consultancy Vanson Bourne in June 2026, 59% of the Brazilian executives surveyed admit that their companies' infrastructure still isn't ready to run AI applications at scale.
The figure matters to whoever signs off on the budget because it shifts the debate. Until now, the AI race was treated as a race for GPUs: whoever bought or rented more computing capacity would win. Dell's survey, with 2,950 respondents across 32 countries and 100 decision-makers in Brazil alone, suggests the real ceiling lies elsewhere.
The bottleneck has moved
For 74% of the Brazilian executives surveyed, storage performance and data access speed already matter more than processing capacity in their companies' critical AI applications. That means adding more chips to the bill doesn't solve the problem: the model may run fast, but it spends much of its time waiting for data to arrive from slow or poorly integrated systems.
For a long time, the discussion about technology capacity focused mainly on processing, but AI applications also depend on fast data access, adequate storage, and integration across different environments.
Joel Brawerman, director of the Infrastructure Solutions Group at Dell Technologies for Latin America
For the founder, the translation into monetary terms is direct: every point of latency in data access turns into idle GPU cost, and every poorly built integration between systems turns into an AI project that never leaves the pilot stage. The question that should be on the budget meeting agenda isn't just "how much processing do we need," but "how much are we paying to wait for data to arrive."
Fragmented infrastructure is costly
The second bottleneck identified by the study is structural: 68% of Brazilian executives say their companies' infrastructure isn't unified. In practice, that means on-premises servers, one or more public clouds, and legacy systems coexisting without full integration, which spreads applications out but complicates exactly the resource AI consumes most: organized, accessible data.
Companies' response has been to consolidate. Among respondents, 86% say they're looking to concentrate servers and storage on fewer, but more powerful, platforms. The stated priorities for modernizing data centers, servers, and storage reveal where the money should go first:
| Priority | % of executives |
|---|---|
| Cybersecurity and resilience | 46% |
| Performance and scalability | 45% |
| Simpler management and automation | 41% |
It's worth noting that security leads the list, ahead of performance. That's compliance becoming an infrastructure decision before the model even runs, not after: whoever decides where the data lives also decides how much regulatory risk the company carries.
The energy bill has reached the boardroom table
The third axis of the study is energy, and here the number is the toughest one for anyone planning an AI expansion: 86% of Brazilian executives consider energy efficiency essential in infrastructure purchasing decisions, and 67% say they will only expand AI use if it can be done without a significant increase in energy consumption.
In short: the barrier to scaling AI is no longer just technical, it has become a recurring operational cost equation. Training and running a model in production consumes electricity continuously, and that cost doesn't show up in the launch headline, it shows up in the following quarter's power bill.
Brazil's backdrop reinforces the urgency of the issue. The country is trying to attract data center investment through Redata, a tax-incentive regime that requires the use of clean energy, and this week the United States began requiring a license to export advanced AI chips to Brazil, which ABES, Brazil's software industry association, already classifies as an unequal competitive condition for Brazilian companies compared to American subsidiaries. For anyone planning AI infrastructure expansion in the country, energy and chip access have stopped being a technical variable and become a geopolitical one.
What the founder should ask the CTO
The survey provides ammunition for a different conversation with the technical team, less about how many GPUs to buy and more about where the data lives and how much it costs to move. Here are some questions the study suggests putting on the board's agenda:
- What is the average data access latency in AI applications already in production, and is it being measured?
- How many different environments (cloud, on-premises servers, legacy systems) store the data the model needs to query?
- If we double our generative AI usage, what is the projected increase in energy consumption, and does the current vendor commit to any cap?
- In the event of an audit or incident, do we know exactly where each piece of data used by the model is stored and under which jurisdiction?
None of these questions requires understanding inference architecture. They require understanding where the risk is, and that's exactly where the infrastructure decision stops being an IT matter and becomes a board matter.
The study's own limitation
A point of honesty worth noting for anyone using this data in an internal presentation: the survey measures perception, not direct infrastructure measurement. The 100 decision-makers surveyed in Brazil answered based on their own assessment of the company they work for, not on an independent technical audit.
That doesn't invalidate the signal, especially since the same pattern shows up in the other 31 countries in the global sample of 2,950 respondents. But it means the 59% figure for unprepared companies is a picture of executive self-assessment, useful for calibrating investment priority, not a technical report on installed capacity. For the founder, the practical lesson is to separate two questions that usually get mixed up in the same meeting: "we think we're ready" and "we've measured that we're ready." Dell's survey only answers the first one, and the second is the one the 2027 budget will demand.
Translated from the Brazilian Portuguese original · Read the original
Anthropic's IPO doesn't prove a new AI playbook, it proves there's only one
Anthropic debuts on the Nasdaq valued at up to $1.4 trillion thanks to API and subscription revenue among programmers. But the $121.3 billion contract with Google reveals that the IPO's real risk is the same one that looms over OpenAI: compute bought before demand is proven.







