AI and lasers: why data centre efficiency matters

AI training has a dirty little secret: a lot of the energy isn’t spent on “thinking” at all. It’s spent on moving data between GPUs fast enough to keep those expensive chips busy. When your cluster gets big, the network becomes the bottleneck—bandwidth, heat, cost, and eventually, physics.

That’s why I paid attention when Enlightra came out of stealth with $15M in funding to build chip-scale multiwavelength lasers for data centres. This isn’t “AI software” news. It’s AI infrastructure news. And for SaaS companies and startups scaling internationally, infrastructure trends matter more than most marketing teams want to admit.

This post is part of our “Tehisintellekt idufirmade ja SaaS-ettevõtete turunduses” series, where we usually talk about AI in content production, multilingual campaigns, and go-to-market execution. Here’s the twist: the next wave of AI marketing capability depends on what happens inside data centres—and the economics of power and interconnects.

Enlightra’s bet: copper is the scaling bottleneck

Answer first: Enlightra is building lasers that help data centres replace power-hungry electrical (copper) connections with denser, more efficient optical links, so AI clusters can scale without energy and cost rising in lockstep.

Most modern AI training involves huge GPU clusters. The bigger the model, the more you need GPUs to communicate constantly—sharing gradients, shuffling data, synchronising updates. When that communication is limited by copper links, you hit a wall: you can buy more GPUs, but you can’t feed them efficiently.

Enlightra’s approach is based on a patented multicolour “comb-laser” platform. Instead of using many separate lasers (each doing one thing), they consolidate many wavelengths into a single integrated laser source. Each wavelength acts like an independent data channel.

A simple way to think about it:

• Copper links: more bandwidth often means more power, more heat, bulkier cabling, and shorter reach.
• Optical links with multiple wavelengths: you can push more data through fibre using parallel channels, often with better energy characteristics and higher density.

Enlightra says it has built 8- and 16-channel lasers aligned with AI interconnect needs, and has demonstrated error-free data transmission at target speeds and power levels. Pilot production is planned for 2027, and the lasers are designed around industry-standard silicon photonics fabrication, which is a big deal for manufacturability.

Why multiwavelength lasers are suddenly a SaaS concern

Answer first: If data centre networking becomes meaningfully more energy-efficient, the cost curve of AI (training and inference) improves—directly affecting SaaS gross margins, pricing power, and how aggressively you can scale AI features.

SaaS leaders often treat infrastructure as “someone else’s problem” (cloud provider, hosting partner, etc.). That’s fine until infrastructure costs start showing up as:

• Higher COGS for AI-heavy products (think summarisation, assistants, personalisation, recommendations)
• Regional scaling friction (adding EU capacity to meet latency or compliance requirements can be expensive)
• Sustainability reporting pressure (especially for B2B SaaS selling to enterprise and public sector)

Here’s what’s changed in 2025: AI isn’t a side feature anymore. Many SaaS companies are shipping AI across the core product, and the usage curve is steep. You don’t just pay for compute; you pay for the data centre reality behind compute—including interconnect.

So when a deeptech startup claims they can help decouple AI performance growth from energy growth, it’s not abstract science. It’s a potential shift in the unit economics of AI-enabled SaaS.

The hidden constraint: GPU-to-GPU communication

A lot of teams budget for “GPU hours” and forget the rest. But at scale, communication overhead can dominate:

• synchronisation across GPUs
• data movement across racks
• networking gear power draw
• cooling overhead triggered by dense interconnect

Even if Enlightra isn’t “the” solution, the direction is clear: AI infrastructure is prioritising optical interconnects because copper is hitting practical limits.

From lasers to marketing: the bridge most teams miss

Answer first: Better AI infrastructure translates into cheaper, faster, and more reliable AI capabilities—which changes what’s realistic in multilingual SaaS marketing and international expansion.

This series focuses on practical AI for startups: content velocity, localisation, and smarter demand gen. But those tactics increasingly rely on AI systems that are cost-sensitive.

When AI costs drop (or even just stabilise), marketing teams can do more of the following without finance shutting it down:

• Always-on personalisation across languages (not just top-of-funnel ads)
• Real-time lead enrichment and scoring using deeper models
• Dynamic website experiences per industry/segment
• High-volume experimentation with creative, landing pages, and messaging

I’ve found that many teams over-invest in “AI tooling” and under-invest in AI operations discipline—cost controls, performance monitoring, model choice, caching strategies, and prompt governance. Infrastructure improvements help, but disciplined teams benefit first.

A concrete example: multilingual campaigns at scale

If you’re an Estonian SaaS expanding to DACH or Nordics, you quickly face a choice:

• translate a small set of pages and run conservative campaigns, or
• run always-on multilingual iteration across ads, landing pages, emails, and in-product onboarding.

The second option wins—because speed compounds. But it also consumes more AI inference (generation, rewriting, summarising, scoring). Lower infrastructure cost and better data centre efficiency make that approach easier to justify.

So yes, lasers can influence your marketing pace. Not directly. Economically.

What to watch next: practical signals for founders and CMOs

Answer first: Track interconnect and energy efficiency innovations as leading indicators of AI cost trends—and translate them into product and marketing decisions early.

Enlightra’s story highlights a broader pattern: data centres are being redesigned around AI workloads. If you’re building or marketing SaaS with AI at the core, these are the signals that matter:

1) Energy efficiency is becoming a product constraint

Enterprises increasingly ask vendors about:

• where data is processed
• energy footprint (even if roughly estimated)
• sustainability posture and reporting

You don’t need perfection, but you do need a narrative grounded in operational reality. If your AI features are expensive and slow, customers notice.

2) Optical interconnect is a scaling theme, not a niche

When startups raise money for things like comb-lasers, it’s a clue that hyperscalers and AI cluster builders are hunting for ways to:

• increase bandwidth density
• reduce power per bit
• fit more compute into the same facility limits

If you sell SaaS internationally, you benefit when cloud providers can scale capacity without power becoming the limiting factor.

3) Hardware innovation will reshape “what’s normal” in AI tooling

The marketing tech landscape tends to follow infrastructure.

As AI becomes cheaper:

• “AI for content” becomes baseline
• differentiation moves to workflow integration and data advantage
• the winners ship features that feel instantaneous, multilingual, and context-aware

The uncomfortable truth: if your SaaS product’s AI still feels like a slow add-on in 2026, users will treat it like one.

Action plan: how SaaS teams can benefit before 2027

Answer first: You don’t need to wait for Enlightra’s pilot production to act; you can improve AI economics now by treating AI usage like a measurable system.

Here are practical moves that work for startups and scale-ups.

1) Build an “AI cost per outcome” dashboard

Stop tracking “tokens” or “GPU spend” in isolation. Track cost against business outputs:

• € per qualified lead
• € per sales-ready account
• € per activated user
• € per retained customer after AI onboarding

This reframes AI from an expense to a controllable growth engine.

2) Design your multilingual stack for reuse

If you’re doing multilingual campaigns (a big theme in this series), optimise for reuse and consistency:

• approved terminology per industry
• reusable message frameworks per persona
• content modules (headlines, proof points, CTAs) that can be recombined

The goal is fewer “from scratch” generations and more guided variation.

3) Use a model portfolio, not a single model

One model for everything is the fastest path to unnecessary cost.

A simple portfolio approach:

• smaller, cheaper model for first drafts and classification
• higher-end model for final copy, strategic pages, or sensitive industries
• deterministic rules for compliance-critical phrasing

4) Cache and batch where it doesn’t hurt UX

If you generate the same summaries, snippets, or recommendations repeatedly, you’re donating margin to your cloud bill.

Common caching wins:

• FAQ answers and help centre drafts
• account-level insights refreshed daily
• lead summaries updated on key events

5) Treat sustainability as a growth narrative, not a checkbox

If you sell to enterprise buyers, you can connect efficiency to credibility:

“We scale AI features without scaling energy waste at the same rate.”

That line only works if your operations back it up—monitoring, optimisation, sensible defaults.

Where this fits in the “AI in marketing” series

Enlightra’s $15M raise is a reminder that AI progress isn’t only about better prompts or better models. It’s also about the physical systems that keep AI affordable.

For SaaS and idufirmad, the practical takeaway is simple: AI-driven marketing and international expansion get easier when AI infrastructure gets cheaper and more efficient. Optical interconnect, silicon photonics, and multiwavelength lasers sound far from your day-to-day campaigns—until they show up as lower inference costs, faster response times, and fewer constraints on experimentation.

If you’re planning your 2026 growth roadmap, here’s the question I’d keep on the table: Which parts of your marketing and product experience would you scale 10× if AI costs dropped by half—and are you building the processes to take advantage of that moment?