Conca, Lfg Zest: in AI, every dollar buys fewer gigawatts and 2027 becomes decisive

The artificial intelligence cycle cannot be understood simply by adding up the investments announced by the major cloud players. For Alberto Conca of Lfg Zest, the crucial step is to understand what those capital outlays are actually buying and which companies are able to retain a growing share of them. In the October 14 statement, his thesis is that a sharp rise in spending can be accompanied by a much more limited increase in the electrical power associated with new infrastructure. That would not prove that AI has stopped growing, but rather signal a change in the composition of growth and in the distribution of benefits along the supply chain.

In the scenario outlined by Conca, between 2026 and 2028 capital devoted to AI infrastructure would rise by about 46%, while added gigawatts would increase from 15.1 to 16.8, up by about 11%. These are estimates from the analysis, not actual figures for the period, and their meaning must be read precisely: gigawatts express electrical power, not the number of calculations or the quality of the service produced. More advanced systems can deliver different performance levels at the same power level. The comparison Conca proposes therefore concerns above all how much capital is absorbed for each unit of energy infrastructure, not a full measure of AI productivity.

System costs change the way investments are read

Conca’s explanation starts with the shift between technology generations. Within the scope considered by his model, one gigawatt of systems based on Nvidia Grace Blackwell would correspond to roughly $46 billion, while with Vera Rubin it would exceed $60 billion. These are not universal price lists, since configurations and included items may differ, but rather the order of magnitude used to describe a more expensive and denser technology. In an infrastructure constrained by available power, the rise in technological content can therefore absorb more capital even without a proportional increase in installed gigawatts.

From this perspective, the thesis is not that every new investment produces less utility, but that monetary growth alone does not make it possible to infer how much the physical size of the market is expanding. Conca estimates that the value captured by Nvidia per gigawatt could rise from about $25 billion with Blackwell to $40 billion with Vera Rubin. The component maker could thus grow revenue even if infrastructure expands more slowly. The question for the investor becomes which factor is supporting the results: new deployments, more content per facility, or higher prices.

Memory is the most price-sensitive point

For Conca, memory makes the gap between revenue growth and volume growth particularly clear. The statement cites for Micron a sequential revenue increase of 74% against an estimated rise in the amount of memory sold of around 4%. The first figure refers to the third fiscal quarter of 2026, not the latest quarter available in October; the second remains Conca’s own analytical estimate. Micron’s official communications for that quarter describe, in DRAM, a modest increase in bits sold and strong price growth, together with a favorable product mix effect.

To update the timing reference, Micron’s results published on September 30 report fourth-quarter fiscal revenue of $54.23 billion, versus $41.46 billion in the previous quarter, equal to about 30.8% more. This later figure cannot be automatically compared with the 4% cited by Conca, because the period and scope must match. The update also does not resolve the issue that matters to the analyst: when a significant share of growth depends on prices and higher-value products, revenue alone does not show how much physical supply has increased or how lasting margins are.

2028 is a normalization risk, not a certain deadline

The risk identified by Conca arises from the meeting point between new memory production capacity and a possible slowdown in the expansion of AI infrastructure. In his view, part of the new plants would make an impact mainly from 2028 onward; if demand were to become less explosive precisely then, suppliers’ pricing power could weaken. It is a conditional scenario: the launch of new plants does not by itself guarantee rebalancing, because demand growth, the technology mix and actual production timing also matter. In its official comments on the third quarter, Micron foresaw a gradual improvement in supply in 2028 without saying when it would meet demand.

The distinction is important so that a cyclical hypothesis is not turned into a mechanical forecast. Conca is not suggesting that solid balance sheets today should deteriorate immediately: he is highlighting the possibility that still-very-strong results may coexist with less favorable future conditions. The risk for a stock valuation emerges when exceptional prices and sustained margins are projected for too long. In this view, the resilience of the AI theme is not enough to make all the companies exposed to it equivalent.

Industrial companies do not all participate in the same way

His review of the industrial supply chain includes 239 companies, of which 41 would have significant exposure to data centers and only five would derive more than 40% of revenue from the theme. These numbers describe the sample in the analysis, not the entire universe of suppliers. Conca nevertheless notes that the market tends to group many companies into the same narrative, that of AI infrastructure expansion, bringing expectations and multiples closer together even when revenue sources differ. For the analyst, exposure to the theme can become very visible in valuations without translating in the same way into margins.

The most useful distinction is between equipment manufacturers and companies that carry out the work. The former can benefit from scarcity, specialization and pricing power, but remain exposed to normalization when supply catches up. Contractors, in Conca’s reconstruction, tend instead to grow volumes more while keeping margins relatively stable, and can be affected above all by weaker new orders. That is why the signal to watch would be found not only in the income statement, but in new orders and in the backlog still to be executed: indicators capable of anticipating future revenue.

Why Conca is looking at 2027 orders

2027 becomes central because the orders collected during that year will feed part of the activity in 2028. Within the industrial spending framework considered by Conca, the growth rate could fall from about 46% to about 6%. The statement does not fully spell out the time base of the first rate, so the comparison should be kept as the author’s scenario without reconstructing an annual series that was not provided. The economic point is nonetheless clear: 6% growth still indicates an increase in spending, not a 40% reduction in its level. Nor should it be confused with the different cumulative 2026–2028 comparison used at the beginning of the analysis.

Conca also notes that, in his framework, the announced investments support 2026 consensus, while for 2027 and 2028 expectations would require about 12–15% additional capital compared with what is already committed. That is a gap to watch, not an automatic demonstration that the cycle is unsustainable. The conclusion of his analysis is therefore selective: it is necessary to distinguish between those who depend on added gigawatts, those who sell more technology per gigawatt and those who benefit from temporarily exceptional prices. The next phase of AI, in his thesis, will be read in the quality and duration of these revenue sources, rather than in the mere total of the billions announced.