Sponsored by
Our Founder Was Early to AI Winners. Now He’s Tracking What Comes Next.
Over the last 18 months, MavSource’s founder tracked and invested around major AI-related names like Micron +100%, Nvidia +74%, Sandisk +130%, Western Digital +74%, TSM +22%, Broadcom +27%, Okta +35% and Lam Research +39% — an average return of approximately +63% across the group.
Now, he’s bringing the sources and ideas behind those insights to you in one daily email digest. MavSource aggregates the most important AI updates from newsletters, podcasts, company news, AI labs, funding rounds, and more — then summarizes what matters in a simple 5-minute brief.
Past performance does not indicate future results. Informational only; not investment advice.
If you still haven't downloaded my "Subscribe and Download your Prompt guide book today from Above and Unlock 200+ AI-Powered Income Ideas" book...
AI Spotlight · The Future of Efficient AI
Green AI: Why Your Favorite Tools Are Going Lean
The biggest AI models in the world are quietly going on a diet. The reason is not the environment. It is survival, and it is reshaping which companies win the next decade.
For the past three years, the dominant story in AI has been scale. Bigger models. Larger datasets. More compute. More power. The assumption baked into nearly every major AI lab's strategy was simple: if you want smarter AI, you build bigger AI. And for a long time, that assumption held.
In 2026, that assumption is cracking. The cost of running large models has become a genuine crisis for enterprise budgets. Data center energy demand is on track to consume more electricity than entire countries. And a set of techniques that researchers called "Green AI" for years without much fanfare has suddenly moved to the center of the most important conversation in technology: how do you build AI that is fast, cheap, and sustainable all at once?
This issue is a plain-English explainer on what Green AI actually means, why model efficiency has gone from academic curiosity to competitive weapon, and what the techniques behind it, pruning, quantization, and knowledge distillation, are doing to the tools you use every day. No PhD required.
|
What is Green AI? Green AI is the practice of building and running AI systems that achieve the same or better performance while using dramatically less energy, memory, and compute. It is not about making AI worse in the name of sustainability. It is about making AI smarter by making it leaner. The goal is to eliminate the computational waste that comes with brute-force scaling and replace it with precision engineering that gets the same job done with a fraction of the resources. |
* * *
The problem with bigger
AI is quietly consuming power at a national scale
The numbers are not abstract. AI data centers consumed an estimated 53 to 76 terawatt-hours of electricity in 2024. By 2028, that figure is projected to rise to between 165 and 326 terawatt-hours per year in the United States alone, roughly 12 percent of total anticipated U.S. power demand. Globally, data centers are on track to consume close to 945 terawatt-hours by 2030, approaching 3 percent of all electricity used worldwide.
To give that context: if AI were a country, its annual electricity consumption would already rank it among the top 12 most energy-intensive nations on the planet. By 2026, AI data centers are projected to consume more electricity than Japan uses in a year. And this is before agentic AI, which uses exponentially more tokens per task, reaches mainstream enterprise adoption.
|
The hidden cost nobody talks about Generating 1,000 AI images produces carbon emissions comparable to driving a gasoline vehicle for 4.1 miles. At scale, this is not a rounding error. Training a single large language model can consume more energy than five cars produce over their entire lifetimes. The compute cost of training GPT-4 class models is estimated to have used millions of kilowatt-hours. Water usage is the overlooked dimension. Data centers use enormous volumes of water for cooling. Microsoft, Google, and Amazon have all reported rising water consumption figures tied directly to AI workload growth, with combined carbon and water footprints that are now material disclosures in sustainability reports. |
The energy crisis in AI is not coming. It is here. And it is creating an urgent market demand for a different kind of intelligence: not bigger, but leaner. Not louder, but more precise. That is where Green AI comes in.
* * *
The techniques
Three tools making AI dramatically more efficient
Green AI is not a single technology. It is a collection of engineering techniques that work together to strip out computational fat without losing the intelligence underneath. The three most important ones in 2026 are pruning, quantization, and knowledge distillation. Here is what each one actually does, in plain language.
|
Technique 1: Pruning The art of removing what the model does not actually need Think of a large language model as a vast neural network with billions of connections, weights, and neurons all firing together every time it generates a response. The counterintuitive discovery that powers pruning is this: most of those connections are not doing very much. A significant portion of a model's weights contribute almost nothing to its final answers. Pruning is the process of identifying those redundant connections and removing them. The results are dramatic: a well-pruned ResNet-50 model can match its baseline accuracy on standard benchmarks even at 90% sparsity, meaning 90 percent of the model's weights have been set to zero and eliminated. The model shrinks. The cost to run it drops. The quality barely moves.
|
|
Technique 2: Quantization Storing numbers less precisely, without losing the answer Every number inside an AI model is stored with a certain level of precision. By default, most modern models use 32-bit floating point numbers, which represent values with high precision but also consume a lot of memory and computing power. Quantization asks a simple question: do we actually need that much precision? The answer, in most cases, is no. Switching from 32-bit to 8-bit integer precision cuts memory usage by 75 percent, dramatically reduces the computing resources needed for each inference, and lowers energy consumption significantly, all while maintaining accuracy that is indistinguishable in most real-world tasks. More aggressive quantization to 4-bit or even 2-bit is now being tested, with surprisingly small drops in output quality.
|
|
Technique 3: Knowledge Distillation Teaching a small model to think like a large one Knowledge distillation is the most elegant of the three techniques. The process starts with a large, highly capable model, called the teacher. That teacher is then used to train a much smaller, faster model, called the student. The student learns not just from raw training data, but from the teacher's probability distributions, its nuanced sense of which answers are likely and which are not. The result is a small model that reasons far more like its large teacher than you would expect given its size. DistilBERT, one of the most cited examples, is 40 percent smaller than BERT, runs 60 percent faster, and retains 97 percent of its accuracy on language benchmarks. Google's "Distilling Step-by-Step" research showed that small distilled models can outperform their much larger teachers on specific tasks, using a fraction of the training data.
|
Used individually, each of these techniques produces meaningful efficiency gains. Used together in a compression pipeline, they can reduce model size by up to 90 percent without significant accuracy loss, making it feasible to run powerful AI on smartphones, edge devices, and budget infrastructure that no frontier model could ever reach.
* * *
Who is winning
Efficiency is becoming the new moat
The competitive landscape in AI shifted in early 2025 when DeepSeek, a Chinese AI lab, released models that matched GPT-4 class performance at a fraction of the training and inference cost. The response from the industry was not dismissal. It was alarm. DeepSeek demonstrated that efficiency innovations could close the gap between a small team with constrained compute and a frontier lab spending billions.
By April 2026, DeepSeek V4-Flash was available at $0.14 per million input tokens, compared to GPT-5.5 at $5.00. That is a 35x cost difference for comparable coding and agentic tasks. DeepSeek V4-Pro, their long-context model, runs at $0.435 per million tokens, roughly 11 times cheaper than GPT-5.5. These numbers are not theoretical. Teams are actively switching workloads based on them.
|
Three companies betting their future on lean AI Apple has made on-device AI the centerpiece of its 2026 strategy. At WWDC 2026, Apple doubled down on running AI directly on iPhones, iPads, and Macs rather than routing requests to cloud servers. This is only possible because of aggressive model compression. Apple Intelligence features use highly distilled, quantized models small enough to run on device hardware while preserving the quality users expect from frontier cloud models. DeepSeek has turned efficiency into a price weapon. Its Mixture-of-Experts architecture activates only the subset of the model needed for each specific task rather than running the entire model for every query. DeepSeek's sparse attention system, released in its V3.2 experimental model, allows long-context operations at API costs reduced by as much as half, by selectively loading only the tokens that matter for each response. The AI model compression market as a whole is responding. Valued at $1.12 billion in 2025, it is projected to reach $3.58 billion by 2034, growing at a CAGR of 11.8 percent. The market that barely existed three years ago is now one of the fastest-growing segments in enterprise AI infrastructure. |
The pattern is clear: efficiency is no longer a secondary concern that engineers think about after building a model. It is a primary design requirement that determines whether a product can be deployed at all, priced competitively, and run profitably at scale. The companies that master lean AI are the ones that will define the next phase of this industry.
* * *
Beyond model design
Carbon-aware computing: running AI when the grid is cleanest
Green AI goes beyond making models smaller. One of the most practical and underappreciated techniques is carbon-aware computing: scheduling AI workloads to run at times and in locations where the electrical grid is drawing from the highest proportion of renewable energy. The grid is not equally clean at all times. Solar peaks at midday. Wind peaks at night or during storms. Carbon-aware scheduling shifts non-urgent AI training jobs to match the grid's cleanest moments.
|
What carbon-aware computing looks like in practice Time-shifting: Model training runs that are not time-sensitive get automatically scheduled to run during off-peak renewable hours rather than during peak demand periods. The output is identical. The carbon footprint is substantially reduced. Geography-shifting: Computation gets routed to data centers in regions where the local grid happens to be running on higher renewable proportions at that moment. A query processed in Iceland during peak hydropower output carries a fraction of the carbon cost of the same query processed in a coal-heavy grid region at peak demand. Model routing: Different tasks get assigned to models of appropriate size. A simple summarization task goes to a small, efficient model. A complex multi-step reasoning task goes to a larger one. The overhead of using a massive frontier model for simple queries is eliminated. This alone is estimated to yield cost savings of up to 10 times for companies that implement it systematically. |
MIT researchers working on this space have noted that carbon-aware strategies, combined with efficient hardware selection, can dramatically reduce the environmental footprint of AI workloads without any change to the models themselves. The efficiency gains come from when and where you run the compute, not just how you build it.
* * *
What it means for you
The tools you use daily are already being rebuilt around this
You may not see the words "pruning" or "quantization" in any product announcement. But you are already living with their effects. Every time an AI feature runs locally on your iPhone instead of sending your data to a server, that is model compression at work. Every time a faster, cheaper model tier appears in an app you use, efficiency engineering is what made it possible. And every time AI appears in a context where it previously could not fit, an edge device, a browser plugin, a wearable, it is because lean AI made it small enough to get there.
|
Four things Green AI is already changing in products you use Faster responses: Compressed models run inference dramatically faster than their full-size counterparts. The speed improvement you have noticed in AI assistants over the past year is not just better hardware. It is smaller, better-engineered models running the same tasks more efficiently. Lower prices: The dramatic cost reductions in AI API pricing over 2025 and 2026 are largely driven by efficiency engineering. Providers passing on the savings from more efficient models is what has made AI tools cheaper for both businesses and individual users. On-device AI: Features that process your data locally, without it leaving your device, are only possible because distilled and quantized models can fit into the memory constraints of a smartphone or laptop chip. Privacy-preserving AI and efficient AI are, in practice, the same thing. More capable free tiers: When models cost less to run, AI companies can afford to offer more to free users. The generosity of free-tier AI tools in 2026 compared to 2023 is a direct downstream effect of efficiency research that has cut the cost of serving each query. |
The three techniques at a glance
| Technique | What it does | Real-world result |
|---|---|---|
| Pruning | Removes unused weights and connections | Up to 90% model size reduction with minimal accuracy loss |
| Quantization | Reduces numerical precision of weights | 75% memory reduction (32-bit to 8-bit), faster inference |
| Distillation | Small model learns from large teacher model | 97% accuracy at 40% smaller size, 60% faster speed |
| Carbon-aware | Schedules compute to cleaner grid windows | Same output, substantially lower carbon footprint |
* * *
The real story underneath Green AI
The narrative around AI and the environment has mostly been one of alarm: AI is consuming energy at an unprecedented rate and the trajectory is unsustainable. That alarm is legitimate. But there is a second story running in parallel that gets far less attention: the engineers building AI are acutely aware of the problem and are making genuine, measurable progress on solving it from the inside.
Efficiency is not a nice-to-have in 2026. It is a survival requirement. The companies that figure out how to build smaller, faster, cheaper models that deliver the same intelligence as their bloated predecessors are not just being responsible. They are building the competitive advantage that will matter most when energy costs, regulatory pressure, and customer demand for faster AI all converge at once.
The next generation of AI will not be defined by which lab has the biggest model. It will be defined by which lab has the most efficient one. The race is already underway, and lean is winning.
|
The future of AI is not about which model is the biggest. It is about which model can do the most with the least. That race has already started, and the winner will not be who you expect. |
* * *
|
Before you go Have you noticed AI tools in your daily life getting faster or cheaper lately? Or are you already routing different tasks to different models based on cost? Hit reply and tell me what you are seeing. The most interesting reader observations will shape a follow-up issue on how individuals and teams are using lean AI to cut costs without cutting quality. |
Until next time,
AI Spotlight
Practical AI, translated into real work, once a week.