VSORA chair Sandra Rivera tells Bloomberg that AI inference demand is durable but consolidation is coming

Sandra Rivera, Chair of VSORA, appeared on Bloomberg Television's The Close on 24 July 2026, fielding questions on whether AI infrastructure spending can be justified by actual revenue.^[1] The interview is notable less for any product announcement and more for what it signals about where the industry's centre of gravity is moving.

Demand is real, but the bar is rising

Rivera's core message was measured: enterprises are still pulling AI into their workflows to drive efficiency and innovation, and she sees no slowdown in that pull.^[1] At the same time, she stopped short of treating demand as a blank check. A Bloomberg data point published the same month offered context: global AI sales, excluding China, reached $25 billion for hyperscalers and neoclouds in Q1 2026, exceeding the industry's estimated $21 billion in depreciation costs for the second consecutive quarter - a thin but real margin of sustainability.

Rivera's argument on the Bloomberg segment was that two things can be true simultaneously: the frontier should keep advancing, and a regulatory framework that reassures the public is both necessary and compatible with growth.^[1]

The training-to-inference crossover

The more pointed claim in the interview was about where compute investment is heading. Rivera confirmed that the industry is nearing a decisive shift from training to inference.^[1] Her reasoning:

• Frontier-model training remains the province of a handful of players - an extraordinarily expensive and complex endeavor^[1]
• The bulk of AI activity will come from enterprises, organisations, and individuals running models against their own data and objectives
• That workload is inference, and it is where VSORA has staked its architecture^[1]

This framing is consistent with VSORA's own market projections. AI inference is projected to grow at a 16 % CAGR from $124 billion in 2025 to $255 billion in 2030, according to figures the company has cited publicly.

Capital is available, but differentiation is the filter

Asked how a startup competes when advanced silicon is a bottleneck, Rivera was candid: capital is not the constraint.^[1] AI is the top destination for investment, and getting in front of investors is not the hard part. The challenge is demonstrating a differentiated architecture that can survive contact with real production workloads.

VSORA's answer to that test is Jotunn8, its data-centre AI inference processor. The chip delivers 3,200 Tflops of compute power with 288 GB of HBM3e memory, manufactured on TSMC's 5 nm process using CoWoS multi-chip packaging.^[1] The company claims it keeps compute utilisation above 50 % on real inference workloads - against figures of 10 % or less that it attributes to conventional GPUs on the same memory-bound tasks. Those claims are the company's own; independent benchmarks have not yet been published.

Sovereignty as tailwind, not guarantee

Bloomberg's Katie Greifeld pressed on an obvious tension: if VSORA, like nearly every other fabless company, relies on TSMC for manufacturing and packaging, where does European sovereignty actually reside?^[1] Rivera's answer leaned on supply-chain resiliency rather than autarky. Drawing on her years managing geopolitical risk at Intel, she pointed to manufacturing and packaging spreading across regions - including TSMC's own investments in the US and Europe - and cited COVID as proof that no supply chain should trace back to a single geography.^[1] The total addressable market, she noted, is global even when a company starts local.^[1]

The question to watch is whether Jotunn8's production ramp, which began in 2026, produces independently verified performance data before the inference-chip field narrows further. Rivera's Bloomberg appearance raises VSORA's profile at exactly the moment when investors and hyperscaler procurement teams are deciding which second-source suppliers to qualify - and that timing is unlikely to be accidental.