Why Live Learning is the only path to profit
By Dr Yichuan Zhang, CEO & Co-founder, Boltzbit
The model race has become a contest for scarce resources, and that competition is leading the AI industry down an unsustainable path. A handful of companies now control the chips, capital and data centers required to build today’s largest AI models. Alphabet, Amazon, Meta and Microsoft are on track to spend nearly $700 billion on AI infrastructure in 2026 alone, more than triple what they spent just two years ago. A single Nvidia GPU can cost up to $40,000. The cluster of hundreds of thousands needed to train frontier models run into the billions. With data centers consuming as much electricity as a mid-size city, what was once a software industry is starting to look like utility-scale infrastructure.
These staggering capital commitments are not translating to success. A recent MIT report found that 95% of corporate generative AI pilots fail to deliver measurable financial returns.
From scale to sustainability: AGI 2.0 and the race to train smarter
Artificial general intelligence (AGI) 2.0 represents a turning point in how intelligence is built and, critically, where the next investment frontier lies. While tech giants double down on capital-intensive scale, a different trend is gathering momentum. Low-cost open-source models are accelerating, and China is already setting the pace. Alibaba’s Qwen now leads global downloads, user preference and new model adoption. In September 2025, 63% of all new model adoption came from Chinese systems, more than double the US share. This is prompting investors to question whether relentless frontier‑scale spending still earns its keep.
The scientific community is reaching the same conclusion. David Silver, former DeepMind researcher, launched Ineffable Intelligence in late 2025 on the thesis that large language models (LLMs) trained on human data represent a ceiling. His company raised $1.1 billion in April 2026 at a $5.1 billion valuation. Turing Award winner Yann leCun has made an equivalent wager. Having publicly argued for years that autoregressive LLMs are a dead end, LeCun left Meta in late 2025 to found AMI Labs, raising $1 billion at $3.5 billion valuation in March 2026.
Investors are paying attention because the economics are shifting. One of the biggest drivers is training cost: training expenses have risen roughly 2.4× per year since 2016, climbing from tens of thousands to tens of millions of dollars, with projections exceeding $1 billion by 2027. Frontier-scale models are increasingly trapped by their own cost curves.
That’s why markets are moving away from “train-everything” general-purpose systems toward domain-adapted models. The evidence that smart data beats big data is impossible to ignore. In early 2025, Li Fei-Fei’s team at Stanford fined-tuned Alibaba’s Qwen 2.5-32B-Instruct model on a curated dataset at a cost of under $50, achieving performance comparable to top-tier reasoning models like OpenAI’s o1 in mathematics and coding capabilities.
AGI 2.0 goes further: models that learn instantly, continuously adapting rather than being retrained from scratch. This reduces cost, speeds up deployment cycles and opens the door to intelligence that scales with use, not with spend.
Defining the new AI economics
1) Sustainable intelligence creation
When billion-dollar budgets become the barrier to innovation, opportunity collapses. Live learning changes the economics. Experiments with the Boltzmann Learning Machine show that models learning as they are used can cut compute and energy costs by orders of magnitude, matching or surpassing fine-tuned models with up to 22-point accuracy gains for only 10% more runtime and 2GB of additional memory. This means fewer retraining cycles, faster iterations and far more capital-efficient intelligence.
2) Ownership and sovereignty
AI’s most valuable asset isn’t compute, it’s the intelligence created through customer data, workflows and proprietary knowledge. Under AGI 2.0, that intelligence stays with the user. Models learn inside a company’s boundary and deploy in isolated environments, so private intelligence never becomes free fuel for another platform. In a world where data-residency rules already constrain how and where training can happen, sovereignty becomes not just a principle but a competitive advantage.
3) Adaptive and interoperable intelligence
Today’s AI stack is powerful but rigid: layers of tools that don’t learn from each other. AGI 2.0 calls for systems that adapt not just to people but to other models, tools and environments. Interfaces reshape themselves through natural language. Knowledge flows across agents. Live learning replaces retraining. When intelligence adapts in real time, it becomes infrastructure: fluid, interoperable and self-improving.
The counterargument: scale still wins, for now
Scale still wins benchmark tests and pretraining breadth but the economics are tilting. Research has shown that each new generation of frontier models costs exponentially more while delivering smaller functional gains. Meanwhile, BOLT-style experiments demonstrate that live learning at inference delivers higher accuracy with minimal extra compute. Adaptation flips the cost curve: capability improves with use, not capital. Meta’s stock fell sharply after its Q1 2026 earnings report as investors focused on the scale of its AI spending plans, underline the fragility of a strategy built purely on capex.
Where the next billion-dollar category emerges
The biggest risk in AI is not missing the next model. It is losing ownership of the intelligence investors are paying to create. They should demand clarity on three metrics: the share of spend going to retraining, the exposure to third-party data licensing, and the speed with which deployed models adapt. The companies that control their data, keep lineage traceable and adapt models “in flight” will operate on faster cycles, lower unit costs and defensible moats. The next winners in AI won’t be those who spend the most on compute but those who turn adaptation itself into the engine of scale.
