Category: technology | 技術 | 기술 | テクノロジー

2026 AI outlook introduction

Marc Andreessen’s 2026 AI outlook was published by A16z. As one of the leading funder of Silicon Valley startups, his world view matters.

I’ve gone through and contrasted his 2026 AI outlook through the lens of the viewpoint I researched and wrote up. The core point over where we differ: Andreessen see’s the dawn of a new unbound industrial revolution. Whereas I think that the upside he sees is hard to navigate to; and also risk mirroring the echoes of financial bubbles past in many of the high profile pure play AI companies like OpenAI or C3.ai.

Andreessen has been the quintessential techno-optimist for at least the past two decades with his emblematic essay Why Software Is Eating the World. His worldview is of an industry where demand is insatiable and revenue is undeniably real.

Like Andreessen in his 2026 AI outlook, I would agree that players like Alphabet, Amazon and Microsoft are making money to variable degrees from their AI offerings as part of cloud computing and productivity software bundles.

But in my view there are two aspects of concern:

• Pure plays with the notable exception of Anthropic don’t seem to have a clear path to profitability in a time period that would match their implied future revenues based on loans and valuations.
• Hyperscalers like Meta and Microsoft are using unusual partnerships to fund their infrastructure and have been inconsistent over how fast they would depreciate their AI computing hardware.

Both Mr Andreessen and myself agree that we are witnessing a technological shift of seismic proportions, arguably, “bigger than the internet” as he put it.

I think that change will happen slower in the short term due to an economic sand box acting as a rate limiter on infrastructure and derived labour efficiencies. In particular, the economic viability of the infrastructure being built to support it.

The Nature of the Boom: Real Revenue or Irrational Exuberance?

A key question is the the solidity of the current market.

For Andreessen, the AI boom is not a speculation fuelled exercise but a demand-driven reality. He argues that unlike the early internet, which required years of physical infrastructure build-out before finding business models, AI is generating cash immediately.

“This new wave of AI companies is growing revenue like… actual customer revenue, actual demand translated through to dollars showing up in bank accounts at like an absolutely unprecedented takeoff rate.”

Andreessen points to “revealed preferences”, observed insight from what people do is more insightful than what they say when asked by market researchers. While the US and European publics express fear of job losses, they are simultaneously adopting AI tools at a fast pace.

My “Dot LLM Era” report, however, suggests this revenue may be dwarfed by the capital expenditure required to generate it. 

It posits that the sector faces a “self-defeating economic” cycle. The report highlights that the current valuations of the “Magnificent 10” tech giants (including Nvidia, Microsoft, and Alphabet) imply a forward price-to-earnings (P/E) ratio of 35 times. This is alarmingly close to the S&P 500’s P/E ratio at the peak of the dot-com boom, which approached 33.

In the report I warn that current valuations assume LLMs will drive revenue growth by $1–4 trillion in the next two years. 

If this growth is achieved through massive job automation, the resulting unemployment could depress the very economy needed to sustain these companies. Like the 2008 financial crisis, this would invite various forms of regulatory intervention.

This implies a sweet spot between speed of productivity gains versus efficiencies in terms of job losses realised by clients – and acts as a break on the velocity of adoption.

The Infrastructure Trap: The Amortisation Crisis

Perhaps the most critical technical divergence between our viewpoints concerns the hardware powering this revolution.

The “Dot LLM Era” report introduces a chilling concept: Amortisation Risk. It draws a comparison to the “telecoms bubble” of the late 90s, where companies laid massive amounts of fibre optic cable.

• The Telecoms Analogy: Fibre optic cable had a useful life of over a decade, meaning even after the companies went bust (like WorldCom), the infrastructure remained useful for Web 2.0.
• The AI Reality: Modern AI infrastructure relies on GPUs and TPUs. These processors have a useful life of only 3 to 5 years before becoming technically obsolete.

The report argues that if an AI bust occurs, the hardware will not be waiting for a resurgence; it will be electronic waste. 

Hyperscalers are currently lengthening the assumed useful lives of this hardware in their financial filings—Google to 6 years, Meta to 5 years—which the report suggests may artificially overstate profits and adversely affect competitors with debt securitised against AI data centre hardware.

The rapid obsolescence of chips represents a “financial and technological amortisation risk” that could lead to trillions in write-offs, similar to the $180 billion loss left by WorldCom.

Andreessen views this hardware cycle through a different lens: Elasticity.

He acknowledges that massive investment leads to gluts, but argues that “the number one cause of a glut is a shortage”. He believes the price of AI compute is falling “much faster than Moore’s Law”.

As prices collapse, demand will expand exponentially. If chips become cheap and plentiful, AI can be embedded into everything, moving from massive “God models” in data centres down to small models running on local devices.

Geopolitics: The US-China Race

Both Andreessen and I agree that the AI landscape is a bipolar contest between the United States and China, but their assessments of the leaderboard differ.

Andreessen frames this as a “new Cold War” where the US must maintain dominance. He is encouraged by the “DeepSeek moment”, referring to a powerful open-source model released by a Chinese hedge fund. To him, this proves that catching up is possible and that the US cannot rest on its laurels. He advocates for open source as a way to proliferate American standards globally.

I offer a more sobering assessment of American pre-eminence. I argue that unlike the 1990s “Long Boom,” where the US was the undisputed hegemon; the current era is defined by high debt and strong competition. The US maybe on the Soviet side of a Reagan era ‘Space Defence Initative (aka Star Wars)’ AI race from an economic perspective.

• Alibaba’s Qwen model claims to deliver comparable performance to American models while requiring 82% fewer Nvidia processors to run.
• China doesn’t have a electrical power crunch in the same way that western data centres have due to its sustained investment in coal, nuclear, gas and renewable power sources.
• Even Silicon Valley investors and major companies like Airbnb are opting for Chinese open-source models because they are “way more performant and much cheaper”.

Andreessen worries about regulation stifling US innovation, I think that the real geopolitical threats are:

• China’s ability to operate largely immune to US sanctions, smuggling chips and utilising data centres in neutral geographies like Malaysia.
• China’s speed at propagating the use of AI within its own populace, driving utility at a lower cost per token than their US competitors with state regulation defining trial ‘sand pits’. This will drive new uses faster in China and in China’s client states across the global south.

The Outcome: Transformation or “Minsky Moment”?

Where is this all heading? Andreessen is betting on a future where AI becomes as ubiquitous and essential as electricity. He envisions a “pyramid” structure: a few massive “God models” at the top, cascading down to billions of specialised, cheap models running on edge devices. He admits these are “trillion-dollar questions,” but his firm is aggressively investing in every viable strategy.

I think that LLMs will make a sustained technological impact, but it will be limited in velocity by economic boundaries. In the “Dot LLM Era” report, I outlined seven potential scenarios, ranging from total transformation to total collapse. Currently, it assigns a ~95% likelihood to the “Moral Hazard” scenario.

• The Moral Hazard: This scenario posits that major AI players will be considered “too big to fail” due to national security imperatives. Governments will step in with loan guarantees and subsidies to backstop the massive infrastructure debts, effectively nationalising the risk . Rather like the banks during the 2008 financial crisis.
• The Telecoms Bust: With a ~75% likelihood, the report fears a “Minsky Moment”, a sudden market collapse driven by the realisation that cash flows cannot cover the massive debts incurred to build short-lived data centres.

Comparative Summary of Perspectives

Conclusion: The Trillion-Dollar Questions

Marc Andreessen candidly admits that “companies… need to answer these questions and if they get the answers wrong, they’re really in trouble”. His firm’s strategy is to bet on everything: large models, small models, apps, and infrastructure. He is doing this on the assumption that the aggregate wave will lift all boats.

My ‘Dot LLM Era’ report offers a counterweight to this enthusiasm. A bubble decouples technological and financial progress. Technological utility does not always equal investor profit. As I note in ‘Dot LLM Era’, “Bubbles don’t kill technology from moving forwards”. The internet did change everything, but it also wiped out trillions in shareholder value along the way.

The defining question for the next five years is whether the demand for AI can grow fast enough to pay for the hardware before that hardware becomes obsolete.

If Andreessen is right, this elasticity of demand would save the day. If I am right, we may be heading for the most expensive recycling project in human history.

The Productivity Paradox and Society

A fascinating tension exists between Andreessen’s view of societal adoption and my own macroeconomic warnings regarding productivity and labour.

The “Wingman” Economy vs. The Phillips Curve

Andreessen describes a “symbiotic relationship” where AI acts as a productivity multiplier: a “wingman” for doctors, coders, and writers. He argues that higher pricing in SaaS can actually benefit the customer by funding better R&D, suggesting a cycle of value creation.

I argue that the wingman sweetspot is optimal but tricky to land, in the report I showed the risk through a darker macroeconomic “thought experiment.” It used the Phillips Curve and Okun’s Law to model what happens if Andreessen AI scenario succeeds too well.

• The Thought Experiment: If AI automation generates $1 trillion in cost savings through job cuts, it implies approximately 10.5 million unemployed US workers.
• The Consequence: Such a spike in unemployment could trigger deflation and a massive drop in GDP. This is “self-defeating economics”: the hyperscalers need a healthy economy to consume their services, yet their success might undermine the investor, enterprise customer and consumer base.

Andreessen counters this fear by citing historical context. He notes the “Committee for the Triple Revolution” in 1964, warned Lyndon B. Johnson that automation would ruin the economy—a prediction that proved false. He believes AI will follow the path of electricity or the internet: initially terrifying, eventually indispensable.

Automation did displace a massive amount of developed world jobs moving at a much slower pace than Andreessen predicted for AI. Electricity moved at an equally slow pace compared to the pace envisioned by AI’s champions.

The Open Source Debate

Both of us agree that the role of open source is pivotal in both narratives.

For Andreessen, open source is the great accelerator. He marvels at how knowledge is proliferating: “Some of the best AI people in the world are like 22, 23, 24”. He views the leak of knowledge as inevitable and beneficial for US competitiveness, provided the US stays ahead.

I analysed the “Red Hat Analogue.” It suggests that in the dot-com era, open-source (Linux) won, but the companies that built the models (or distributions) mostly failed, with Red Hat being the notable exception.

I assigned a ~70-80% likelihood to the “Red Hat Model,” where pure-play LLM creators (like OpenAI or Anthropic) might struggle to justify their capital burn as open-source models like Meta’s Llama or Alibaba’s Qwen commoditise the intelligence.

We have already seen Singapore’s national AI programme drop Llama for Alibaba’s Qwen, reinforcing the idea that the value might accrue to those who service the models, not those who create them.

Final Thoughts

The divergence between Marc Andreessen and my own analysis is not about whether AI works both of us would agree that the technology can be magical and transformative. The disagreement is about who pays for it, how that affects the velocity of AI and who profits.

Andreessen sees a future of abundance where falling prices drive infinite demand.

My own view sees a future of financial reckoning shaping the 2026 AI outlook where shorter hardware lifespans and brutal competition erode margins, setting a slower pace at which we reach Andreessen’s abundance. 

Andreessen’s viewpoint reminded me a lot of mid-20th century aspirations for nuclear power. Nuclear power offered a similar vision in the mid-20th century of electricity too cheap to meter. That was never close to being achieved in the likes of France – arguably the most passionate adopter.

As with the railway mania of the 1840s or the optical fibre boom of the 1990s, society may inherit a significant infrastructure, with a shorter lifespan built on the ashes of investor capital.

Our differing views boil down to a question for the 2026 AI outlook: are we in the “boom” phase, or are we staring down the barrel of the “amortisation crisis”?

As Andreessen himself concluded, “These are trillion-dollar questions, not answers”.