The money flows¶
Understanding where money actually goes in the technology industry requires distinguishing between the impressive announcements and the quieter movements of capital that reveal what investors genuinely believe versus what they say at conferences. The hype flows freely because it costs nothing and occasionally attracts additional investment. The actual money flows more cautiously toward opportunities that might plausibly generate returns, which is a considerably more selective criterion than “sounds impressive during pitch meetings.”
Venture capital has discovered that putting “AI” in a startup’s description increases valuation by amounts that would make traditional financial analysts question their career choices. This has predictable effects on startup naming conventions, pitch deck contents, and the number of companies that suddenly realised their product was AI-powered all along despite previously describing it as sophisticated algorithms or advanced analytics. The investors are complicit in this because admitting that “AI” is sometimes marketing rather than meaningful technical distinction would undermine valuations they’ve already committed to.
The serious money, meaning the amounts that move markets rather than merely funding another cohort of optimistic founders, is flowing primarily into infrastructure. Cloud providers building data centres, semiconductor manufacturers expanding fabrication capacity, and networking companies supplying connectivity represent capital deployment measured in tens of billions of euros annually. This infrastructure spending is both enabling and betting on continued technology growth, which is either prescient positioning or expensive overcapacity depending on whether demand materialises as projected.
Meanwhile, acquisitions are happening quietly as established companies buy promising startups before they become expensive competitors or embarrassing failures. IPO markets remain selective, accepting only companies with plausible paths to profitability rather than the “growth at any cost” approaches that were fashionable during previous enthusiasm cycles. The public markets have remembered that revenue eventually needs to exceed expenses, which is constraining for companies that were planning to rely on perpetual funding availability.
Venture capital’s AI obsession¶
Venture capital investment in AI startups reached record levels in recent years, with funding rounds that would have seemed absurd a decade ago becoming routine for companies with impressive demonstrations and ambitious promises. Seed rounds of €5-10 million, Series A rounds of €30-50 million, and Series B rounds exceeding €100 million are common for AI startups that have products but not necessarily revenue, customers but not necessarily at scale, and technology that works in demonstrations if not always in production.
The valuations reflect expectations about future growth rather than current fundamentals. A company with €5 million in annual revenue might raise at a €500 million valuation, which implies that investors expect revenue to grow by factors of hundreds while maintaining margins that justify the valuation. These expectations are sometimes met, which encourages similar investments elsewhere, but more often are quietly revised downward during subsequent funding rounds that receive less publicity than the optimistic initial announcements.
Foundation model companies, building large language models or equivalent capabilities, attract the largest investments because they require enormous capital for compute infrastructure and training runs before generating revenue. Anthropic, OpenAI, and similar companies have raised billions of euros from investors betting that foundation models will become essential infrastructure that generates corresponding revenue. These investments require believing that AI capabilities will continue improving, that the companies will successfully commercialise their technology, and that they won’t be disrupted by competitors with better technology or lower costs.
Application layer startups, building products using foundation models from others, receive smaller investments but still substantial by historical standards. These companies have faster paths to revenue because they’re not funding expensive model training, but face questions about defensibility if their core technology comes from external providers who might offer competing products. The investment thesis requires believing that application-specific features, data, or customer relationships provide sufficient moats against competition.
Infrastructure and tooling companies, providing development tools, deployment platforms, or optimisation services for AI, receive funding based on the assumption that regardless of which AI applications succeed, infrastructure will be needed universally. This is the “selling shovels during gold rushes” approach, which historically has been more reliable than prospecting directly. The challenge is that infrastructure markets are competitive and often consolidate around a few winners, so being an infrastructure provider is only profitable if you’re one of the survivors.
The venture capital enthusiasm for AI is self-reinforcing in ways that don’t obviously end well. Success stories justify additional investment, which attracts more startups, which increases competition, which pressures everyone to raise more capital to outlast competitors, which increases total capital deployed into a sector where only some participants will generate returns justifying the investments. This pattern is familiar from previous technology cycles and typically resolves through a mixture of consolidation, quiet failures, and recalibration of expectations.
The infrastructure build-out¶
The largest capital deployments in technology are infrastructure investments by established companies building data centres, fabrication facilities, and networking capacity. These investments are less visible than venture funding but involve considerably more capital and reveal what companies with established businesses believe about future demand.
Cloud providers are collectively investing over €100 billion annually in data centre expansion, server purchases, and networking infrastructure. Amazon Web Services, Microsoft Azure, and Google Cloud are building facilities globally to support growing demand for computing, storage, and AI workloads. These investments have multi-year payback periods and are premised on sustained demand growth that might or might not materialise at projected rates.
Semiconductor manufacturers are expanding fabrication capacity with investments that dwarf cloud infrastructure spending. TSMC’s capital expenditure exceeds €30 billion annually for building new fabrication facilities, installing advanced equipment, and ramping production. Intel, Samsung, and other manufacturers are making similar investments, collectively committing hundreds of billions of euros to semiconductor capacity expansion over the next decade. These investments are bets on continued semiconductor demand growth across AI, computing, automotive, and industrial applications.
Memory manufacturers are ramping HBM production with investments measured in billions of euros per manufacturer. SK Hynix, Samsung, and Micron are building production capacity for high-bandwidth memory that AI accelerators require. The investments are substantial but smaller than logic semiconductor capacity because HBM is more specialised and serves narrower markets than general-purpose memory.
Networking infrastructure providers are supplying equipment for AI data centres with different requirements than traditional computing. Connecting thousands of GPUs requires networking with lower latency, higher bandwidth, and different topologies than conventional data centre networks. Nvidia, Broadcom, and others are developing AI-specific networking that commands premium prices and represents substantial capital deployment by data centre operators.
Power and cooling infrastructure represents additional capital requirements that are often overlooked in technology discussions. Data centres consuming tens to hundreds of megawatts require electrical substations, distribution equipment, backup generators, and cooling systems that cost tens to hundreds of millions of euros per facility. These investments are necessary prerequisites for operating the computing equipment but don’t generate revenue directly.
The infrastructure spending is either prescient positioning for sustained growth or systematic overbuilding that will result in excess capacity when growth disappoints. The difference depends on whether AI workloads, cloud adoption, and semiconductor demand continue growing at rates justifying current investments. The infrastructure has long lifecycles, so building capacity today commits to assumptions about demand years into the future when conditions might differ substantially from present expectations.
Acquisitions and strategic investments¶
Established technology companies are acquiring startups both to gain technology and to prevent competitors from acquiring the same capabilities. These acquisitions are less visible than venture funding or IPOs but represent substantial capital deployment and reveal what established companies consider strategically important.
AI startups with promising technology but uncertain business models are attractive acquisition targets because established companies can integrate the technology into existing products and customer bases. Microsoft’s investments in OpenAI, Google’s acquisition of DeepMind (earlier, but relevant precedent), and numerous smaller acquisitions demonstrate that large companies prefer buying promising AI capabilities rather than building everything internally.
The acquisition prices reflect both the technology value and the competitive dynamics of preventing rivals from acquiring the same capabilities. A startup might be worth €100 million based on its current business but sell for €500 million because multiple acquirers are competing and because the acquirer values keeping the technology away from competitors as much as acquiring it themselves. This creates lucrative exits for founders and investors but also means acquirers are paying premiums that might not be justifiable purely on financial returns.
Acquihires, where companies are acquired primarily for their talent rather than technology or business, remain common in AI because skilled researchers and engineers are scarce. An established company might pay €5-10 million per key person to acquire a team, which is expensive by traditional standards but economical compared to recruiting and retaining equivalent talent through normal hiring. The acquired company’s products often get shut down as the team is redirected to acquirer priorities.
Strategic investments, where established companies invest in startups without acquiring them completely, allow the investor to gain exposure to promising technology while maintaining optionality about full acquisition later. These investments provide funding to startups while giving investors board seats, strategic input, and often preferred terms for future acquisition. The arrangements benefit both parties but can create conflicts when the startup’s independent interests diverge from the investor’s strategic preferences.
The acquisition market is selective, favouring startups with defensible technology, proven teams, and strategic fit with potential acquirers. Startups that raised at high valuations but haven’t demonstrated corresponding growth face difficult choices between down rounds, selling at disappointing prices, or attempting independence despite challenging funding environments. The market for mediocre startups is poor, which is creating pressure on the extensive cohort of AI companies funded during peak enthusiasm but struggling to differentiate or achieve profitability.
The IPO market reality check¶
Public markets have been considerably more sceptical of technology companies than private markets, which creates tension between venture-backed companies hoping for lucrative IPOs and public investors requiring profitability or credible paths thereto. The IPO market for technology companies has been selective, accepting only companies with strong fundamentals rather than impressive growth metrics accompanied by spectacular losses.
AI companies attempting IPOs face particular scrutiny because public investors remember previous technology enthusiasms that generated impressive private valuations but disappointing public performance. Companies must demonstrate not just technology capabilities but sustainable business models, reasonable margins, and growth that doesn’t require perpetual cash consumption. These requirements eliminate most AI startups from IPO consideration until they mature substantially beyond their current state.
The few successful technology IPOs recently have been companies with established businesses, profitability, and growth in mature markets rather than speculative bets on emerging technology. This sets expectations that other companies hoping for IPOs must meet similar standards, which is challenging for businesses optimised for growth during abundant private funding rather than profitability required for public markets.
The IPO window opens and closes based on public market appetite for risk, which varies with economic conditions, interest rates, and recent performance of comparable companies. When public markets are enthusiastic, mediocre companies can IPO successfully. When public markets are cautious, only exceptional companies can access public capital. The current environment is more cautious than enthusiastic, which constrains IPO opportunities regardless of company quality.
SPACs (special purpose acquisition companies) provided alternative paths to public markets during peak enthusiasm but have largely fallen out of favour after many performed poorly. The SPAC approach allowed companies to go public with less scrutiny than traditional IPOs, which predictably resulted in many marginal companies accessing public capital and subsequently disappointing investors. The regulatory and reputational damage has made SPACs less attractive for quality companies and less available generally.
The realistic outlook for AI companies hoping for IPOs is that they need to demonstrate considerably more maturity, profitability, and business sustainability than was required during previous enthusiasm cycles. This is healthier for public markets but challenging for companies that raised private capital at valuations predicated on rapid IPO exits rather than long-term private company operation.
Where the serious money actually goes¶
Following capital flows rather than listening to pitch decks reveals that the serious money is concentrated in infrastructure, established companies, and proven business models rather than speculative bets on emerging technology. This doesn’t mean emerging technology receives no funding, but the capital deployed into infrastructure is orders of magnitude larger than venture capital regardless of which receives more media attention.
Cloud provider capital expenditure exceeds the entire venture capital industry’s annual deployment multiple times over. Amazon, Microsoft, and Google each spend tens of billions of euros annually on infrastructure, which is more than all AI venture capital combined. This infrastructure spending is premised on sustained demand for cloud services and AI workloads, which represents a much larger bet on technology trends than any individual venture investment.
Semiconductor manufacturer investments dwarf cloud infrastructure spending. TSMC alone invests over €30 billion annually, and the industry collectively commits hundreds of billions to capacity expansion. These investments are necessary for supplying chips that enable all downstream technology, but they’re bets on continued semiconductor demand growth that might prove optimistic if technology spending slows or if applications require less advanced chips than projected.
Research and development spending by established technology companies represents another enormous capital deployment often overlooked in favour of more visible venture funding. Alphabet, Microsoft, Meta, and Amazon each spend €20-40 billion annually on R&D, much of it directed toward AI and related capabilities. This spending exceeds the entire venture capital deployed into AI startups and represents established companies’ bets on future technology directions.
Share buybacks and dividends by profitable technology companies return over €100 billion annually to shareholders, which is capital that could theoretically be invested in growth but is instead returned because management believes shareholders can deploy it more effectively than the companies can internally. This suggests that even companies with enormous resources and expertise can’t find sufficient attractive investments to absorb all available capital.
The pattern is that infrastructure providers building capacity for others to use receive the largest capital deployments, established companies with proven businesses receive substantial ongoing investment, and speculative ventures receive enough funding to be interesting but not enough to meaningfully compete with established players’ resources. This is rational capital allocation that favours relatively safe bets over speculative ones, but it means that genuinely disruptive innovations must either come from established companies with resources or from startups that can succeed despite resource disadvantages.
The hype versus reality gap¶
The gap between what receives attention and what receives capital is substantial and revealing. Hype concentrates on consumer-facing AI applications, chatbots, and impressive demonstrations that generate social media engagement. Capital concentrates on infrastructure, enterprise software, and business models with clear revenue potential that don’t generate viral videos but do generate cash flows.
Foundation models receive enormous attention because they’re impressive and because their developers are skilled at generating publicity. The actual capital deployed into foundation models, while substantial by startup standards, is modest compared to the infrastructure required to support them. The cloud providers hosting the training runs, the semiconductor manufacturers producing the chips, and the memory manufacturers supplying HBM are capturing more total capital even if receiving less attention.
Consumer AI applications generate enthusiasm and user engagement but struggle with business models that justify venture capital valuations. Many impressive AI demonstrations are loss-making at scale because inference costs exceed revenue per user. Until inference efficiency improves or business models evolve beyond advertising and subscriptions, consumer AI remains more impressive than profitable despite capturing disproportionate attention.
Enterprise AI applications receive less hype but more actual deployment because businesses will pay for capabilities that improve productivity or reduce costs. The use cases are boring compared to consumer applications but the business models are clearer and the customers are willing to pay, which attracts capital from investors who prioritise returns over excitement.
Infrastructure and tooling receive minimal hype because data centre construction and development tools are not inherently exciting. However, infrastructure receives enormous capital deployment because it’s necessary regardless of which applications succeed. The companies building infrastructure are making calculated bets on aggregate demand growth rather than specific application success.
The hype-reality gap exists because attention and capital are allocated by different mechanisms toward different goals. Media attention goes to whatever generates engagement, which favours impressive demonstrations and bold predictions. Capital goes to whatever generates returns, which favours proven business models and necessary infrastructure. Understanding technology markets requires tracking both but weighting capital flows more heavily than hype for predicting what actually gets built versus what gets discussed.
What this means for observers and participants¶
Following the money provides more reliable signals about technology trajectories than following the hype, but requires effort because capital flows are less visible than press releases. Large infrastructure investments by established companies signal confidence in sustained demand growth. Selective venture capital deployment signals investor caution despite public enthusiasm. IPO market challenges signal that profitability requirements are reasserting after a period where growth at any cost was acceptable.
For startups, the implications are that raising capital requires demonstrating clearer paths to profitability than during peak enthusiasm, that infrastructure and enterprise applications attract more reliable funding than consumer applications, and that competition for capital is increasing as funding availability normalises after the post-pandemic abundance.
For established companies, the implications are that infrastructure investments commit to multi-year assumptions about demand growth, that competitive dynamics require matching rivals’ capital deployments regardless of internal confidence in returns, and that acquiring promising startups before they become expensive competitors is increasingly attractive compared to building equivalent capabilities internally.
For investors, the implications are that venture returns from current AI investments will likely disappoint because too much capital chased too few genuine opportunities at too high valuations, that infrastructure plays provide exposure to AI growth without betting on specific applications, and that public markets will remain selective until profitability improves across the technology sector.
The money flows reveal an industry simultaneously confident enough to deploy enormous capital into infrastructure and cautious enough to require increasingly strong fundamentals before committing to speculative ventures. This is either rational calibration between opportunity and risk or collective confusion about whether current enthusiasm is justified by fundamentals. The answer won’t be clear until years from now when we can evaluate whether infrastructure investments generated returns, whether AI startups justified their valuations, and whether the technology lived up to its promise sufficiently to justify the capital deployed in its pursuit. Until then, watching where money actually goes rather than where attention flows provides the most reliable signal available for understanding what sophisticated parties genuinely believe versus what they say they believe when cameras are present.