- AI companies face high compute costs and energy demands while developing and deploying models.
- A crop of startups are building solutions to make AI more energy and cost-efficient.
- From data center cooling to chip efficiency, VCs say these startups are making AI cheaper and greener.
DeepSeek sent a clear message to Silicon Valley startups: It's possible to do more with less.
Now, competitors in the US are scrambling to replicate the Chinese startup's approach, which appears to rival the performance of top AI models in the US but seemingly at a fraction of the cost.
That's prompted industry insiders to question the billions of dollars being spent on AI infrastructure. It has also spotlighted startups that are developing solutions to lower the high costs of developing, deploying, and running AI models.
Training AI requires huge processing power, which is fueled by clusters of graphics processing units, or GPUS. They consume a lot of energy, and these power circuits are largely provided by data centers.
"The energy consumption of training a large AI model can produce emissions equivalent to the lifetime emissions of multiple cars," Andreas Riegler, general partner at APEX Ventures, told Business Insider. "As models grow in size, the demand for energy scales exponentially, raising sustainability concerns for future applications," he added.
There are many approaches toward making AI greener and cheaper to use, Riegler told BI. Startups can improve software efficiency, develop more energy-efficient chips, and tap into renewable energy sources, he said.
BI spoke to seven investors based in Europe and the US, asking them to put forward startups that are helping make AI cheaper and greener to use.
Mobius Labs
Total raised: $7 million
Backers: Atlantic Labs, APEX Ventures, Lunar, and angels
Suggested by: Andreas Riegler, founding partner at APEX Ventures
What it does and why it's one to watch: Berlin-based Mobius Labs has developed an AI platform that claims to reduce AI costs by using 10 times less computing power.
While it launched in 2020, the startup pioneered a new product line in 2024 — which allows it to "transition from a computer vision recognition company to a full multimodal generative AI provider," said the startup's cofounder and CEO, Appu Shaji.
Mobius Labs' platform aims to reduce the energy required to run AI models, which is otherwise costly and energy-intensive.
The startup says it has "quantized" Meta's Llama70B model — essentially, divided it into smaller, more measurable increments — "without losing accuracy, and enabling it to run on a single GPU instead of four," said Riegler, whose VC firm has backed the company. "This method skips unnecessary calculations without sacrificing the accuracy" of the output, he added.
Gemesys
Total raised: $12.8 million
Backers: Atlantic Labs, NRW.Bank, and Plug and Play Tech Center
Suggested by: Andreas Riegler, founding partner at APEX Ventures
What it does and why it's one to watch: AI-hardware startup Gemesys has developed a chip that works similarly to the human brain and positions itself as a more energy-efficient alternative for AI applications.
Launched in 2023, the German startup has designed a piece of hardware called a neuromorphic chip. These differ from traditional chips, which process information in a "step-by-step manner," said Riegler.
"Neuromorphic chips are built to work like a network of neurons (the brain's nerve cells), which can handle multiple things at once and adapt over time," he told BI. "This makes them very good for tasks like pattern recognition, image processing, and sensory data analysis."
According to Riegler, this makes it a faster alternative to traditional chips because it powers AI applications that need to run continuously — such as smart sensors, robots, or small devices — more efficiently.
Cohere
Total raised: $975 million
Backers: Thomvest Ventures, Index Ventures, Tiger Global, Radical Ventures
Suggested by: Umesh Padval, managing director at Thomvest Ventures
What it does and why it's one to watch: Cohere is a foundational model layer company that provides large language models to enterprises so they can develop and rapidly deploy AI applications.
Based in Toronto, Canada, the company allows enterprises to host Cohere's proprietary language model on its own servers. The startup has secured partnerships with the likes of McKinsey and software giant Oracle. In September 2023, it launched a chatbot called Coral.
"The company is providing high-performance LLM models, with the highest accuracy at the lowest cost — thus providing the lowest cost of ownership for its customers," said Padval, an investor in the company.
Cohere raised $500 million in July 2024 to grow its technical team and double down on its enterprise offerings.
Corintis
Total raised: $9 million
Backers: Blueyard Capital, Founderful, and Acequia Capital
Suggested by: Antonia Albert, principal at Founderful, and Rick Hao, partner at Speedinvest
What it does and why it's one to watch: Lausanne, Switzerland-based Corintis has developed a cooling solution for AI applications and data centers.
The startup has created a technology called a "precision microfluidic cooling solution," which aims to improve semiconductor chips' performance by preventing them from overheating.
Corintis embeds a network of microscopic channels directly into the chips. The cooling liquid extracts heat, achieving 10 times more efficiency than current solutions, said Albert.
Apheros
Total raised: $1.85 million
Backers: Founderful
Suggested by: Antonia Albert, principal at Founderful
What it does and why it's one to watch: ETH Zurich spinout Apheros launched in August 2023 with a patented technology that helps data centers cool down.
"We take a device called a heat sink, whose purpose is to bring heat away from your chip, your GPU, and into the surroundings and the coolant — and we've optimized it through the usage of a metal foam," cofounder and CEO Julia Carpenter previously told BI. This process allows heat to dissipate more easily in data centers, ultimately using up less energy in the process.
"By 2030, an estimated 6% of global energy consumption will be used for cooling data centers — and a need for cost-and energy efficient liquid-based solutions is inevitable," said Albert.
Syntiant
Total raised: $121 million
Backers: M12, Microsoft, Intel Capital, Alumni Ventures
Suggested by: Samir Kumar, general partner at Touring Capital
What it does and why it's one to watch: California-based Syntiant has developed a method to make semiconductor chips more efficient by "combining purpose-built silicon with an edge-optimized data platform," said Kumar.
The company says it has streamlined the conversion of raw and synthetic data into "quality machine learning models" — making data processing for AI more efficient.
Kumar told BI that Syntiant is focused on improving power and performance efficiency for inference — a technique in which AI makes predictions about new data — that takes place on battery-powered internet-connected devices. Improved efficiency on these devices could lead to more AI-enabled sensors for speech, sound, vision, and other modalities, he added.
DBtune
Total raised: $3 million
Backers: 42 Cap, Axeleo Capital, JVH Ventures, Tiny Supercomputer, Exxeta Ventures, Tola Capital
Suggested by: Sheila Gulati, managing director at Tola Capital
What it does and why it's one to watch: Sweden-based DBtune is a spinout from Stanford University. Launched in 2022, the startup aims to improve the way databases perform to reduce cloud costs for clients while lessening their environmental impact.
The startup uses AI and machine learning to optimize each database to the specific customer's workload and needs, which, it claims, makes database tuning more efficient.
"The company's core technology is a fully automated service designed to optimize database management, enabling businesses to achieve high database performance while minimizing their environmental impact," said Gulati. "Database efficiency is a game changer for a company's carbon footprints."
Cartesia
Total raised: $26.7 million
Backers: AWS Startups, 515 Ventures, Conviction Partners, Index Ventures, Lightspeed
Suggested by: Sheila Gulati, managing director at Tola Capital
What it does and why it's one to watch: San Francisco-based Cartesia, which was founded in 2023, is developing a new architecture for AI systems.
The startup has released a slate of new AI models that it says can run large models on smaller devices using less energy.
Cartesia "builds real-time, multimodal intelligence for every device using their pioneered State Space Models, enabling more efficient AI," Gulati told BI.
Mako
Total raised: $1.4 million
Backers: Flybridge, Jeff Dean, the Chief Scientist of Google Deepmind and head of Google Research
Suggested by: Chip Hazard, general partner at Flybridge Capital
What it does and why it's one to watch: Previously known as A2 Labs, Mako emerged from stealth in October 2024 with a $1.4 million pre-seed round.
The Boston-based startup aims to reduce compute costs by up to 70%. It has developed a technology that automates the process of GPU tuning — something that increases the processing power of a GPU — which is an otherwise manual and costly task.
Mako's platform generates a code that runs on any kind of hardware and optimizes GPU performance. This, in turn, ensures that AI models can run more efficiently at peak performance, the company website claims.
"Mako is a company to watch as they allow any AI model to be run on any hardware or cloud platform, and brings hardware independence to AI application developer," said Hazard.
PoroTech
Total raised: $26.1 million
Backers: Speedinvest, Cambridge Enterprise, IQ Capital Partners
Suggested by: Rick Hao, partner at Speedinvest
What it does and why it's one to watch: University of Cambridge spinout PoroTech, which launched in 2018, is developing a technology that makes semiconductor chips more energy-efficient.
It works with gallium nitride (GaN), a semiconductor material that is a better conductor than silicon. The team researched how to engineer GaN so that it can be applied efficiently to semiconductor-based applications.
"The company's groundbreaking work in GaN technology has positioned it at the forefront of innovation, particularly in the development of optical interconnects and optical processors that enhance energy efficiency, communication, and computation at the speed of light," Hao told BI. "This saves a ton of energy for data centers, etc."
Liquid AI
Total raised: $296 million
Backers: AMD Ventures, Bold Capital Partners, Capgemini
Suggested by: Chip Hazard, Flybridge Capital
What it does and why it's one to watch: MIT spinout Liquid AI emerged from stealth in October 2023.
The startup is working to build a general-purpose AI system called a liquid neural network. The aim is for this to use less compute power and energy while operating more transparently than the technology that powers mainstream chatbots and image generators.
Liquid neural networks are a type of architecture that processes data more efficiently than traditional neural networks, as they consistently adapt to new inputs.
The startup's "Liquid Foundation Models" can be applied to everything from detecting fraud to controlling self-driving cars.
"Unlike traditional transformer-based models, whose memory usage and inference time increase with longer input sequences, LFMs maintain near-constant inference time and memory complexity regardless of context length," said Hazard. "So they can process longer sequences without a significant increase in computational resources or energy consumption."
