How AI Companies Can Turn Limited Compute Into an Advantage

For an AI company with limited computing resources, competing with well-funded technology giants can seem like an uphill battle. The largest AI developers can invest billions of dollars in advanced chips, massive training runs and increasingly sophisticated foundation models. Smaller companies rarely have the capital to match that spending—and trying to do so can be a costly strategic mistake.

The opportunity lies in recognizing that customers do not necessarily need the biggest model. They need solutions that solve specific problems reliably, affordably and in ways that fit their businesses.

Members of the Senior Executive AI Think Tank, a curated group of experts specializing in machine learning, generative AI and enterprise AI applications, share their perspectives on how AI companies with limited compute resources can compete in an increasingly crowded market. Drawing on their experience across AI development, enterprise technology and business strategy, they explore the challenges facing resource-constrained companies, the strategic decisions that can shape their competitive position and the opportunities emerging as AI technology evolves.

“The frontier labs optimize for breadth; a focused player wins the last mile they ignore.”

– Egbert von Frankenberg, CEO of Knightfox App Design Ltd.

SHARE IT

Own a Niche Before Expanding

Egbert von Frankenberg, CEO of Knightfox App Design Ltd., recommends concentrating on a narrowly defined market where specialized knowledge matters more than general-purpose capabilities.

“Pick a narrow domain,” von Frankenberg says, pointing to areas such as legal discovery, radiology and contact center quality assurance, “and go deeper than any generalist can be bothered to. Proprietary domain data, expert-labeled datasets and workflow integration beat a bigger model that knows everything shallowly.”

The advantage extends beyond model selection. 

“The frontier labs optimize for breadth; a focused player wins the last mile they ignore,” he says.

Von Frankenberg also advises building the system around the model rather than treating the model as the entire product. Retrieval, data quality, evaluation, guardrails and workflow integration can create meaningful differentiation without requiring a company to train its own foundation model.

“A small team can build a better-engineered system on top of an open or third-party model and outperform a better-funded rival who thinks the model is the product,” he says. “Smaller fine-tuned models, distillation, caching and smart retrieval let you serve a specific job at a fraction of the cost.”

Make User Experience a Competitive Advantage

Edward Morris, CEO and Lead Prompt Engineer at Enigmatica, argues that technical capability alone does not determine which products customers adopt.

“User experience. Focus on that,” Morris says.

He points to ChatGPT’s popularity as evidence that users return to products for reasons beyond raw model performance. 

“ChatGPT is the most popular AI platform there is currently. The reason why people go back to ChatGPT constantly is because of the experience the user has,” he says. “Not because it’s necessarily better.”

For resource-constrained companies, this creates an opportunity to make complex technology feel simple. Understanding users’ daily routines, minimizing unnecessary steps and delivering consistent results can build loyalty without requiring enormous training budgets. A product that fits naturally into existing work can outperform a technically impressive alternative that customers find cumbersome.

“You have to focus on who your users are and tailor for them,” he says.

Solve Expensive Problems, Not Abstract AI Challenges

Fabio Danze Montini, Investor and Owner of FDM Industrial Sales & Marketing SL, emphasizes starting with a customer problem rather than a technology ambition.

“Limited compute is only a disadvantage if you are competing on the model itself,” Montini says. His recommendation is to identify “a painful, valuable and poorly solved business problem” and use the minimum technology required to solve it better than competitors.

“Customers do not buy parameters, benchmarks or GPUs,” he says. “They buy lower cost, faster decisions, more revenue, less risk or less work.”

This approach also changes how companies allocate scarce resources. Instead of investing heavily in capabilities before validating demand, a startup can establish customer relationships, refine its product and develop proprietary knowledge around a clearly defined workflow. That knowledge can become harder for general-purpose competitors to replicate.

Turn Proprietary Data Into a Defensible Moat

Jim Liddle, Entrepreneur, Investor, Advisor and Enterprise AI Strategist, advises startups to compete through positioning rather than infrastructure spending.

“You can’t out muscle large, well-funded AI vendors, so you need to out position them,” Liddle says. His approach centers on combining focus, speed and proprietary data to establish a meaningful advantage in a specific market.

“I mentor startups who often ask this type of question, and my standard answer is that they should look to own a niche deeply,” he says, citing law, manufacturing and medicine as potential areas for model specialization. “Depth beats breadth, especially when you can’t afford breadth.”

Liddle also stresses the importance of developing data that competitors cannot easily access. 

“One proprietary, focused, specialized dataset beats a thousand generic scraped ones,” he says.

Maximize Every Unit of Compute

Vivek Kumkar, Sr. GenAI Leader at Amazon Web Services (AWS), recommends treating compute utilization as a core business discipline rather than an infrastructure afterthought.

“Compete on utilization, not acquisition,” Kumkar says. “A rival can buy 10 times your GPUs and still waste half of them.”

His approach emphasizes batching requests, caching repeated results and selecting appropriately sized models for individual tasks. 

“A small team that runs hot, batching, caching and right-sizing models to the job, pulls more usable compute from a modest fleet than a giant pulls from a huge one,” he says.

Infrastructure that sits idle still incurs costs, he notes. So startups should measure cost per successful task, latency and utilization rather than relying solely on hardware capacity or benchmark scores.

“Their money is locked into long training runs,” he says. “Your edge is inference discipline you sharpen every week, cheaper and faster than capital can chase.”

“Treat the big players’ spending as your subsidy, not your threat. ”

– Mahendran Vasagam, Principal Member of Technical Staff at Slack Technologies

SHARE IT

Stay Model-Neutral and Learn From Every Correction

Mahendran Vasagam, Principal Member of Technical Staff at Slack Technologies, recommends using advances from larger AI developers without becoming dependent on a single provider.

“Treat the big players’ spending as your subsidy, not your threat,” Vasagam says. “Every billion they spend on training lowers the price of the models you can rent.”

He advocates model neutrality, allowing a company to switch between providers as performance, pricing and customer requirements change. 

“A frontier lab has to sell its own model; you can switch to whichever model is best and cheapest each quarter,” he says.

Vasagam also identifies customer corrections as a valuable source of differentiation.

“Every time a user fixes an output in your product, you learn what ‘right’ looks like in their work,” he says. 

Capturing that feedback responsibly can create a continuous improvement loop, turning everyday usage into domain-specific intelligence that external model providers may never see.

Build for the Devices Customers Already Use

Uttam Kumar, Engineering Manager at American Eagle Outfitters, recommends focusing on lightweight, associate-facing AI assistants that work on existing retail hardware.

“Concentrating engineering efforts on developing lightweight associate-facing assistant tools that operate smoothly on existing in-store hardware allows nimble teams to bypass costly cloud infrastructure expenses entirely,” Kumar says.

He emphasizes that practical deployment matters more than theoretical sophistication. Physical retailers need “practical, low-latency workflow solutions right on the sales floor,” he says, especially when those solutions must work with legacy point-of-sale systems.

This strategy can reduce infrastructure demands while addressing an immediate operational need. By designing for the devices and systems customers already have, smaller companies can avoid expensive deployments and deliver value where employees actually work. The broader lesson applies to any industry where reliability, speed and integration outweigh the appeal of a general-purpose AI interface.

“Compute is an advantage, but it should not define competitive strategy.”

– Aishwarya Shah, Independent Researcher

SHARE IT

Measure Success by Business Outcomes

Aishwarya Shah, Independent Researcher, argues that compute should inform an AI company’s strategy without dictating its ambitions.

“Compute is an advantage, but it should not define competitive strategy,” Shah says. Instead, resource-constrained companies should prioritize “specialized applications, smaller optimized models, proprietary domain knowledge and measurable business outcomes.”

She also highlights strategic partnerships as a way to access foundational capabilities without recreating the infrastructure that larger developers have already built.

“The competitive advantage is not merely doing more with less; it is understanding which problems deserve to be solved and delivering trusted, economically sustainable solutions,” Shah says.

Her advice encourages leaders to evaluate AI investments through measurable outcomes such as productivity, customer satisfaction, reliability and cost reduction. Those metrics help companies distinguish genuine differentiation from technological novelty and focus limited engineering resources where they can create the greatest value.

Use Smaller Models and Conventional Software Together

Nikhil Jathar, CTO of AvanSaber Inc., recommends examining the full cost of completing a customer task correctly before choosing a model.

“Choose one expensive customer task and measure the full cost of completing it correctly, including retries and human checks,” Jathar says. Companies can then “remove unnecessary model calls and use smaller models wherever testing supports them.”

Microsoft’s Phi-3 Mini research describes a 3.8-billion-parameter model designed to run on a phone, he adds, demonstrating that useful language capabilities do not always require frontier-scale infrastructure.

“A small company can build on such models and spend its scarce compute on difficult exceptions,” he says.

Jathar also recommends combining AI with deterministic software. For invoice processing, conventional code can verify totals while AI interprets ambiguous text. This division of labor can lower costs without sacrificing reliability.

Move Intelligence Closer to the User

Pawan Anand, Portfolio Head for Communications, Media and Entertainment at LTM, recommends targeting environments where limited connectivity, privacy and response times matter more than access to massive cloud infrastructure.

“A compute-constrained AI company should choose problems where scarce compute is an advantage, not a handicap,” Anand says, highlighting edge computing, on-device AI, private deployments and low-latency applications.

He advises companies to “build for a narrow job, make every token and watt earn its keep, and push intelligence as close to the user or machine as possible.”

This approach creates an edge over large companies “whose economics and architectures assume abundant cloud compute.”

“Constraints can force better product discipline, lower operating costs and a deployment model competitors cannot easily replicate without redesigning their stack,” he says.

10 Strategies for Competing With AI Industry Giants

• Own a narrow market before expanding. Develop deep domain expertise, high-quality proprietary data and workflow integration that general-purpose models may not provide.
• Treat user experience as a core differentiator. Make the product intuitive, reliable and tailored to the specific needs of its users.
• Start with a costly customer problem. Validate demand and demonstrate measurable business value before investing in sophisticated technology.
• Build a proprietary data advantage. Capture permissioned, domain-specific information and feedback that improve performance over time.
• Optimize compute utilization. Measure cost per successful task and improve batching, caching, model sizing and inference efficiency.
• Remain model-neutral. Preserve the flexibility to switch providers as pricing, capabilities and enterprise requirements change.
• Design for existing hardware and workflows. Reduce deployment costs by making AI work with the devices and systems customers already use.
• Tie investments to business outcomes. Evaluate AI initiatives against measurable improvements in cost, productivity, reliability and customer value.
• Combine AI with conventional software. Use deterministic code for predictable tasks and reserve model calls for ambiguity and complex interpretation.
• Explore edge and on-device opportunities. Prioritize use cases where privacy, connectivity, latency and low operating costs reward efficient architectures.

Winning in AI Is About More Than Compute

For AI companies operating on limited compute, the temptation to measure success against the industry’s biggest players is understandable—but it can also be a trap. Chasing the same benchmarks, infrastructure and capabilities risks turning a smaller company’s greatest limitation into the center of its strategy. The more important question is where a company can deliver something its competitors cannot easily replicate, whether that means solving an overlooked industry problem, earning customers’ trust or making AI work where larger systems fall short.

As AI capabilities become more widely accessible, companies that know their customers better, make sharper strategic bets and execute with greater precision will gain the upper hand. Compute still matters, but it is not the only resource that determines who wins. For smaller AI companies, the path to lasting competitiveness begins with recognizing that they do not have to build the most powerful AI in the market—they have to build the AI that matters most to the customers they serve.