With over 20 years of experience, we transform your digital presence. We specialize in website and E-Shop development, SEO and Digital Marketing, ERP software and smart automation that take your business to the next level.
Address:
154 Egnatia Street, Thessaloniki 546 36
Phone:
+30 231 231 4095
E-mail:
hello@twodots.gr
See how real-time AI inference and servers like Lumai Iris Nova are changing personalization, latency and cost in e-shops.
The announcement of the Lumai Iris Nova server, as presented by Design News, is not just another news from the semiconductor and data center space. It is a signal of where the AI market is heading: from the period of impressiveness around big models, we are moving into the period of efficient, fast and economically viable execution. Simply put, the big question for businesses is no longer just «what can AI do?», but «how quickly, at what cost and at what point in the customer experience can it do it?» This is where AI inference comes in, that is, the phase in which an already trained AI model provides answers, predictions or suggestions in real time.
For an e-shop owner, AI inference is closer to everyday operation than it seems. When a visitor sees recommended products, when an AI chatbot answers a question about availability, when a fraud detection system evaluates a payment, or when a dynamic pricing mechanism adjusts prices based on demand and inventory, there is inference behind all of this. Training is the ’building« of intelligence. Inference is the moment when intelligence works in front of the customer and influences the sale. That is why the news about a new AI server like the Lumai Iris Nova has commercial interest: it shows that the market is looking for solutions for lower latency, higher computational performance and lower energy burden, especially in real-time AI inference applications.
Lumai is moving in the field of photonic computing, that is, the use of optical technologies for calculations that are currently mainly done with electronic circuits. Without meaning that every e-commerce business will buy such a server tomorrow, the direction matters. As the use of generative AI, LLM inference and personalized experiences increases, the more pressure traditional GPU infrastructures are under. Thus, the discussion about AI accelerators, GPU alternatives and more efficient data centers is not a technical detail; it is a matter of customer service costs, checkout speed, quality of proposals and ultimately profit margin.
E-commerce has learned to measure everything: conversion rate, average order value, customer acquisition cost, returns, repeat purchases, advertising costs. The next mature metric will be the performance of AI decisions per interaction. If the recommendation engine displays the right suggestion within a fraction of a second, the customer stays in the purchase flow. If the AI chatbot takes several seconds to respond, the user returns to Google or opens the next store. If omnichannel personalization is not synchronized between site, email, social and physical store, the experience becomes disjointed. AI inference, then, is not just backend infrastructure; it is part of the customer experience.
According to McKinsey, 711% of consumers expect personalized interactions from businesses, while 761% are disappointed when they don’t. For an e-commerce store, these percentages translate into very specific decisions: they need e-commerce personalization on the homepage, personalized searches, relevant bundles, intelligent abandoned cart recovery, and better post-purchase support. As the chart below shows, the expectation of personalization is no longer a premium feature; it’s a core element of trust.
Source: McKinsey, The value of getting personalization right
The first application point is product search. A modern e-shop cannot rely only on exact word matching. The user can write «wedding shoes», «comfortable sneakers for travel» or «gift for new dad». AI inference allows for semantic search, understanding intent and matching with products that do not necessarily contain the same words in the title. The second point is product recommendations. A recommendation engine that takes into account browsing behavior, purchase history, profit margin, seasonality and availability can increase the average order value without pressuring the customer with irrelevant upsells.
The third point is service. An AI chatbot that works with access to return policies, inventory, order status, and product specifications can reduce tickets, respond after hours, and unblock purchases that would otherwise be lost. The fourth point is risk prevention: fraud detection, unusual purchases, suspicious addresses, or incompatibilities between payment and shipping details. The fifth is pricing and merchandising, where dynamic pricing and smart campaigns can be based on real demand, available stock, and competitive behavior. In all of this, real-time AI inference is critical, because the decision is only valuable if it comes while the customer is still on the site.
The growth of artificial intelligence has a less visible side: energy consumption and pressure on data centers. The International Energy Agency reports that electricity consumption from data centers, AI and cryptocurrencies was about 460 TWh in 2022 and could exceed 1,000 TWh by 2026. This trend explains why companies like Lumai are investing in new architectures, such as photonic computing, and why the market is looking for any possible AI accelerator that can offer a better performance-to-consumption ratio. For e-shops, the issue is not only environmental. If the cost of inference increases, the AI functions you want to offer at scale become more expensive.
The chart below shows the IEA’s estimate of the increase in electricity demand from data centers, AI, and cryptocurrencies. The image helps us understand why energy efficiency will directly impact the pricing of cloud services and AI tools.
Source: International Energy Agency, Electricity 2024
The second hidden cost is latency. In e-commerce, latency is not a technical metric for the development team; it is a barrier to entry. Google has published that as mobile page load time increases from 1 to 3 seconds, the probability of a bounce increases by 32%, while from 1 to 5 seconds it increases by 90%. Add to this AI functions that are slow to respond, such as search, recommendations or chatbots, and the problem becomes commercial. AI infrastructure should be designed so that it does not add weight to the experience, but makes it more immediate.
As the graph below shows, even small delays can dramatically change user behavior. This is where cloud latency, edge AI, and choosing the right AI server all come together to drive conversion rates.
Source: Think with Google / SOASTA research
There’s a third cost that often goes unnoticed: cart abandonment. The Baymard Institute estimates the average cart abandonment rate at 70.19%. It’s not all down to speed or the absence of AI, of course. It’s down to shipping costs, mandatory account creation, slow checkout, lack of trust, and poor experience. But AI inference can step in at critical moments: displaying clarification on returns, suggesting an alternative payment method, identifying high intent and triggering a relevant offer, or helping the customer complete the purchase without searching for information.
The graph below shows how much room for improvement there is at checkout. Even a small reduction in abandonment can have a bigger impact than an expensive advertising budget increase.
Source: Baymard Institute, Cart Abandonment Rate Statistics
Step 1: Map out where an AI decision can directly impact revenue or cost. Don’t start with the tool; start with the customer flow. Note at what stages the user searches, compares, hesitates, abandons, or asks for support. For each stage, define a potential use case: semantic search in search, recommendation engine on the product page, AI chatbot at checkout, fraud scoring at checkout, dynamic pricing on high-demand products. This will help you avoid the trap of investing in generative AI without a clear commercial outcome.
Step 2: Set measurable goals. For example, reducing support response time by 30%, increasing add-to-cart rate from suggested products, reducing bounces on search pages, improving checkout conversion, or reducing manual tickets. AI inference should be judged by business KPIs, not just technical metrics like tokens per second or model size. Technical metrics matter, but only when they are linked to experience and profitability.
Step 3: Check your data. No model will give correct suggestions if the product feed has errors, if the categories are unclear, if the inventory is not updated or if the product descriptions are poor. Before investing in LLM inference, clean up attributes, sizes, colors, prices, availability, profit margins and transaction histories. The quality of the data is often more decisive than the choice of model.
Step 4: Start with a low-risk pilot. A good place to start is with AI search in a specific category or a chatbot that only answers questions about policies, shipping, and returns. Measure latency, accuracy, fallback rate, conversion impact, and user feedback. If the pilot proves valuable, gradually expand to more complex features like personalized bundles or omnichannel personalization.
Step 5: Decide where inference should be performed. For some functions, the cloud is sufficient. For others, such as very fast search, real-time recommendations, or in-store applications, edge AI or a hybrid architecture may make sense. The news about Lumai Iris Nova shows that the market is experimenting with specialized infrastructures that promise better inference performance. However, for most commercial enterprises, the right question is not «what is the most advanced hardware?», but «what architecture delivers a reliable experience at a sustainable cost?».
Step 6: Set governance rules. The AI chatbot needs to know when to refer to a human. Dynamic pricing needs to have limits so as not to destroy trust. Product recommendations need to consider availability and commercial strategy, not just click-through probability. AI in e-commerce is not an autopilot; it is a decision support system that needs monitoring, policies, and continuous improvement.
The choice of infrastructure should start from four criteria: speed, cost per interaction, agility and data control. Speed is about how quickly the system responds to the user. Cost per interaction is about the real cost of each recommendation, search or chatbot response. Agility is about whether you can change the model, provider or architecture without rebuilding the entire system. Data control is about privacy, compliance and protection of commercial information. In this context, technologies such as photonic computing or new forms of AI accelerator are important because they push the market towards more efficient AI inference, but the business decision must remain practical.
For a small to medium-sized e-shop, the most realistic path is to leverage managed AI services, closely monitor costs, and build internal knowledge around data, prompts, response evaluation, and customer journeys. For a larger marketplace or retailer with high traffic, the issue becomes more complex: custom model serving, caching, vector databases, hybrid cloud, edge nodes, or collaboration with providers that invest in specialized hardware may be needed. There, news like Lumai Iris Nova takes on strategic importance, because it shows which technologies might reduce the cost of real-time AI inference at scale.
The conclusion is clear: the next battle in e-commerce will not be decided just by who has more products or a bigger advertising budget. It will be decided by who can deliver the most relevant, fastest and most reliable experience to every user. AI inference is the engine that turns AI models from impressive demos into everyday commercial decisions. E-commerce owners who will treat it as part of a customer experience strategy, and not as an isolated technology trend, will be better positioned in a market where speed, personalization and operational cost are now becoming competitive advantages.
Design News: Lumai Unveils IRIS Nova Server for Real-Time AI Inference
International Energy Agency: Electricity 2024
McKinsey: The value of getting personalization right
Think with Google: Mobile page speed benchmarks
Baymard Institute: Cart Abandonment Rate Statistics
AI inference is the process where a trained artificial intelligence model provides answers or suggestions in real time. It is critical for e-commerce because it impacts the customer experience, offering personalized suggestions and improved support.
The announcement of Lumai Iris Nova demonstrates the market shift towards lower latency and energy-efficient solutions. These technologies improve real-time AI inference, making it more affordable and cost-effective for enterprises.
Key applications include product discovery with semantic search, product recommendations, customer service via chatbots, and risk prevention such as fraud detection. These features improve the overall user experience.
Cost is affected by energy consumption, latency and the infrastructure used. New technologies such as photonic computing offer solutions for more efficient use of resources, reducing costs.
An e-shop can use managed AI services, strictly monitor costs and focus on improving its data. The right choice of infrastructure and the gradual implementation of AI solutions will improve the customer experience at a sustainable cost.
Speed and personalization directly impact conversion rates and customer satisfaction. Consumers expect fast and relevant interactions, and AI inference helps achieve these goals.
Hidden costs include power consumption and latency, which can impact pricing and user experience. Proper management of these parameters is critical to success.
