The state of the realm¶
The technology landscape currently resembles Ankh-Morpork during one of its periodic enthusiasms, where everyone is simultaneously convinced that enormous fortunes are to be made, that the current situation cannot possibly continue, and that they personally must act immediately or be left behind. These three beliefs are not entirely compatible, but that has never stopped anyone from holding them with absolute conviction while investing other people’s money accordingly.
Artificial intelligence has achieved the status of a transformative technology while simultaneously remaining mostly experimental. Large language models can generate plausible text, write code, and convince investors to write cheques for billions of euros, which are admittedly impressive capabilities even if not quite the artificial general intelligence that press releases sometimes imply. The technology works, within limitations that vendors discuss quietly and customers discover loudly, and the question occupying boardrooms is whether this represents the beginning of genuine transformation or merely an expensive interlude before everyone returns to doing things the old way because it was cheaper and more reliable.
The infrastructure supporting this AI enthusiasm is straining in ways that suggest either rapid expansion or imminent collapse depending on your temperament and investment position. Supply chains for critical components are concentrated in geographically precarious locations. Power consumption is approaching levels that make environmental commitments difficult to square with electrical bills. Cloud providers are building data centres faster than they can reliably fill them with paying customers while simultaneously facing hardware shortages that prevent building data centres fast enough. These contradictions are presumably resolvable but not obviously so.
Meanwhile, quantum computing continues its long tradition of being five to ten years from practical utility regardless of which year you ask. Blockchain has largely subsided into specialised applications after discovering that most problems don’t actually benefit from distributed trustless consensus. Edge computing exists in the perpetual state of being obviously necessary but not quite compelling enough for universal deployment. The metaverse is discussed less frequently now, which is probably for the best.
Where the money is flowing¶
Venture capital is pouring into AI startups with the enthusiasm of prospectors who’ve heard there’s gold in them hills but haven’t personally verified the reports. Funding rounds for companies with impressive technology demonstrations but uncertain business models reach valuations that would make traditional investors reach for the smelling salts. The logic appears to be that AI is transformative, therefore companies doing AI things will be valuable, therefore investing now positions you to profit later when the transformation becomes clear. This logic is either prescient recognition of paradigm shift or expensive confusion about the difference between impressive technology and sustainable business models.
The established technology companies are investing even more heavily, which is either validating the AI enthusiasm or demonstrating that nobody wants to be the executive who failed to invest in the next big thing regardless of whether it was actually the next big thing. Google, Microsoft, Meta, and Amazon are spending tens of billions of euros annually on AI infrastructure, research, and development. These investments are large enough that they materially affect these companies’ profitability while being justified by projections about future revenue that are necessarily speculative given that many AI applications are still being discovered.
Hardware manufacturers are enjoying unprecedented demand. Nvidia’s market capitalisation has increased by hundreds of billions of euros based largely on demand for AI accelerators. TSMC is building fabrication facilities at record pace to meet semiconductor demand. Memory manufacturers are ramping HBM production. Power infrastructure suppliers are receiving orders for data centre electrical systems at scales previously associated with industrial facilities. The entire supply chain is experiencing boom conditions that will either prove justified by sustained demand or result in expensive overcapacity when the cycle turns.
Cloud computing revenue growth has reaccelerated after a post-pandemic slowdown, driven partly by AI workloads and partly by continued migration from on-premises infrastructure. The cloud providers are investing their substantial cash flows into expanding capacity, which requires judgement about demand trajectories years into the future. They’re essentially betting that computing will continue moving to centralised cloud providers rather than being disrupted by edge computing, on-premises repatriation, or some other architectural shift that makes current data centre expansion look premature.
The money is also flowing into less visible infrastructure. Networking equipment for connecting AI accelerators, cooling systems for managing heat loads, power distribution equipment for delivering megawatts reliably, and facilities construction for housing everything. The AI boom is creating demand throughout the technology supply chain, which is excellent for suppliers and concerning for anyone worried about whether all this infrastructure will find productive use once deployed.
The capability plateau question¶
AI capabilities have improved dramatically over the past few years, but whether this improvement continues at similar rates is uncertain. Large language models have progressed from GPT-3 to GPT-4 to various competitors with broadly comparable capabilities. Each generation is better than the last, but the improvements are increasingly incremental rather than revolutionary. The curve might be continuing exponential improvement, or it might be approaching diminishing returns. Nobody knows for certain, but substantial capital is being deployed based on assumptions about which it is.
The scaling hypothesis, which states that larger models trained on more data with more computation will continue improving predictably, has held remarkably well so far. This is why companies keep training ever-larger models consuming ever-more resources. The hypothesis might continue holding, or we might discover fundamental limitations that prevent indefinite scaling. Early signs could go either way depending on how optimistically you interpret recent results.
Architecture innovations provide an alternative to pure scaling. Attention mechanisms, mixture of experts, retrieval augmentation, and other techniques improve model capabilities without simply making everything bigger. These innovations are valuable but don’t necessarily change the fundamental trajectory of whether AI is approaching transformative capability or hitting capability ceilings that limit practical applications.
Multimodal models that handle text, images, audio, and video represent another direction of improvement. These models demonstrate broader capabilities than text-only predecessors but face questions about whether multimodal understanding represents genuine progress toward general intelligence or merely expansion of narrow capabilities across more domains. The philosophical debate is interesting; the practical impact is that multimodal models enable new applications while requiring even more computational resources.
The plateau question matters economically because if capabilities are still improving rapidly, then current infrastructure investments are justified by future capabilities that will enable valuable new applications. If capabilities are plateauing, then current applications must justify the investments without counting on dramatic future improvements. The difference is between building infrastructure for a growing market and overbuilding for a market that’s already approaching saturation.
Geopolitical technology fragmentation¶
The comfortable assumption that technology would remain globally integrated is collapsing under geopolitical tensions. The United States and China are increasingly decoupling their technology ecosystems through export controls, investment restrictions, and strategic competition. Europe is attempting to chart a middle course while building technological sovereignty that reduces dependence on both American and Chinese technology. The result is a fragmenting global technology landscape that’s less efficient but possibly more resilient.
Semiconductor restrictions are the most visible manifestation. The United States restricts exports of advanced chipmaking equipment to China, limits Chinese access to cutting-edge semiconductors, and pressures allies to implement similar restrictions. China is investing heavily in domestic semiconductor capability to reduce dependence on foreign suppliers. These efforts will eventually produce multiple semiconductor ecosystems with different capabilities and serving different markets, which is less efficient than the previous global integration but reflects strategic priorities overtaking economic optimisation.
AI development is fragmenting similarly. Chinese AI companies have limited access to advanced Nvidia GPUs due to export restrictions, which drives development of domestic alternatives and potentially creates divergent AI development paths. The United States and allies are concerned about Chinese access to AI capabilities with potential military applications. China is concerned about dependence on Western technology that could be disrupted through geopolitical tensions. Both sides are investing in domestic capabilities that reduce interdependence.
Data sovereignty and privacy regulations are creating regulatory fragmentation. Europe’s GDPR, China’s data protection laws, and various national regulations create different requirements for data handling, which forces companies to either comply with multiple regulatory regimes or segregate operations by geography. Cloud providers build regional data centres to satisfy residency requirements. Technology companies maintain separate data handling procedures for different jurisdictions. The global internet is fragmenting into regional variations with different rules and capabilities.
Standards battles are emerging where geopolitical blocs promote competing technical standards for everything from 5G networking to AI governance. These battles are partly about technical merit and partly about strategic positioning because controlling standards provides advantage in future technology development. The result is duplication of standards development effort and potential incompatibility between systems using different standards.
The fragmentation increases costs, reduces efficiency, and complicates global technology deployment. It also provides resilience against single points of failure and reduces vulnerability to disruptions in any one region. Whether this trade-off is worthwhile depends on your assessment of geopolitical risks versus economic efficiency, which is inherently judgemental and likely to be evaluated differently by different parties with different risk exposures.
Sustainability tensions¶
Technology companies announce ambitious sustainability commitments while deploying infrastructure that consumes electricity at unprecedented rates. These positions are not obviously compatible, which creates tensions that are managed through creative accounting, renewable energy procurement, and optimistic projections about future efficiency improvements.
Data centre power consumption is growing rapidly driven by AI workloads, which is problematic for companies claiming to be carbon neutral. The standard response is purchasing renewable energy credits that theoretically offset fossil fuel consumption, which is legitimate accounting but doesn’t change the instantaneous load on electrical grids that are substantially powered by fossil fuels. The renewable energy being credited often would have been built anyway, which makes the additionality questionable even if the accounting is technically correct.
Water consumption for cooling is creating conflicts in water-stressed regions. Data centres consume millions of litres daily, which is modest compared to agriculture or residential use but still significant when concentrated in specific locations. Some facilities use treated wastewater or seawater, which reduces conflict with other uses but introduces additional operational complexity. The water consumption is increasingly facing regulatory scrutiny and community opposition in drought-affected regions.
Electronic waste from continuously refreshing hardware is substantial. AI accelerators depreciate rapidly as new generations offer substantially better performance, which drives frequent hardware replacement. The replaced hardware has some remaining utility but is often retired because it’s no longer economically optimal, which creates electronic waste streams that recycling infrastructure struggles to handle sustainably. The waste problem is acknowledged but not actively solved because the economics favour replacement over prolonged use.
The sustainability tensions are managed rather than resolved. Companies purchase renewable energy, invest in efficiency improvements, and make public commitments about future carbon neutrality. These efforts are real and meaningful but don’t eliminate the fundamental tension between growing computational workloads and environmental constraints. The situation will likely persist with companies doing enough to manage reputation risk while not fundamentally changing trajectories that are driven by business imperatives rather than environmental concerns.
The talent shortage that isn’t¶
Technology companies complain endlessly about talent shortages while simultaneously conducting mass layoffs, which suggests that the shortage is either narrowly specific or substantially exaggerated. The reality is that demand for certain specialised skills exceeds supply while demand for general technical skills fluctuates with business cycles and funding availability.
AI expertise is genuinely scarce relative to demand. Researchers with deep understanding of machine learning, experience training large models, and publication records at top conferences can command extraordinary compensation because they’re wanted by everyone building frontier AI systems. This scarcity is real and drives compensation inflation that’s spreading to adjacent skillsets as companies discover that anyone who can plausibly claim AI expertise can negotiate aggressively.
The definition of AI expertise is expanding to include people with increasingly tangential qualifications, which suggests that the scarcity is being addressed through credential inflation rather than genuine skill development. Someone who took an online course in machine learning is marketed as an AI specialist. A data scientist becomes an AI researcher through title change. This grade inflation makes assessing actual capabilities difficult and increases the risk of disappointed expectations when AI experts turn out to be moderately skilled practitioners with impressive titles.
General software engineering skills are abundant relative to demand, which is why tech layoffs are possible without companies collapsing from inability to staff projects. The reality is that most technology work is relatively straightforward application development that many engineers can perform adequately. The talent shortage narrative serves to justify high compensation and H1-B visa programmes but doesn’t reflect the actual supply-demand balance for general skills.
The talent situation varies dramatically by geography. Silicon Valley and other technology hubs have deep talent pools but correspondingly high compensation expectations and living costs. Remote work has distributed talent more broadly, which helps companies access skills outside traditional hubs but also means competition for talent is less geographically constrained. The shift to remote work has been partially reversed through return-to-office mandates, which suggests that employers value in-person collaboration more than access to geographically distributed talent or that they’re attempting to reduce headcount through attrition rather than layoffs.
Where things might be heading¶
Predicting technology trajectories is reliably humbling because surprises arrive regularly and confident predictions age poorly. That said, certain directions seem plausible based on current trajectories and underlying drivers, with the usual caveats about uncertainty and unexpected developments.
AI will likely continue improving but at diminishing rates as low-hanging fruit is exhausted and remaining improvements require progressively more resources. The current generation of large language models will be refined, deployed more widely, and integrated into more applications. Revolutionary breakthroughs might occur but cannot be predicted reliably, so planning should assume incremental progress rather than discontinuous leaps.
Infrastructure buildout will continue until either demand saturates or financial constraints impose discipline. The current enthusiasm for building AI data centres will persist while funding is available and while companies fear being left behind. Eventually, utilisation rates and return on investment will matter, at which point infrastructure growth will moderate to match actual demand rather than projected demand based on optimistic scenarios.
Regulation will increase as governments respond to AI capabilities, privacy concerns, competition issues, and geopolitical considerations. The regulatory trajectory is toward more constraints on data collection, algorithmic transparency requirements, and competition enforcement. Technology companies will adapt to regulation as they always have, through compliance where necessary and lobbying to shape rules favourably where possible.
Consolidation will accelerate in mature technology markets while new markets remain fragmented during early growth phases. Cloud computing, social media, and e-commerce are largely consolidated. AI is currently fragmented with many well-funded startups but will likely consolidate as capital requirements increase and as it becomes clear which approaches work commercially. This is the standard technology industry pattern and no reason to expect AI will differ.
Technology will remain central to economic activity and innovation regardless of whether specific trends like AI live up to their hype. Computing becomes more powerful, communication becomes faster, and automation handles progressively more tasks. These trends are decades old and will likely continue because they’re driven by fundamental economic incentives rather than temporary enthusiasms. Individual technologies might disappoint but the broad trajectory of increasing technological capability seems robust.
The current moment is characterised by unusual uncertainty about which technologies will prove transformative and which will prove expensive distractions. AI might revolutionise knowledge work or might prove to be sophisticated automation of routine tasks without the broader capabilities enthusiasts predict. Quantum computing might eventually become practical or might remain permanently five years away. The metaverse might emerge as an important platform or might be remembered as an expensive mistake. These uncertainties are what make the current technology landscape simultaneously exciting for optimists and concerning for anyone responsible for deploying capital or making strategic decisions based on technology trends that remain genuinely uncertain despite confident pronouncements from interested parties.
The Patrician’s perspective would be that uncertainty is permanent, that grand predictions should be discounted heavily, and that survival requires preparing for multiple scenarios rather than betting everything on any single vision of the future. This approach lacks the inspirational quality that makes for compelling conference keynotes but has the considerable advantage of occasionally being correct about the messiness of reality compared to the tidiness of PowerPoint presentations.