As demand for artificial intelligence accelerates, so too does concern over the energy intensity of AI datacentres. In response, Terakraft, a Norway-based renewable-powered datacentre operator, and Neurophos, a next-generation AI hardware startup, have announced a collaboration aimed at delivering sustainable, high-performance AI compute.
Together, the two companies seek to prove that combining green datacentres with breakthrough optical chips can chart a path toward scalable, ultra-efficient AI infrastructure—balancing exponential compute needs with environmental responsibility.
The energy dilemma of AI
AI has rapidly shifted from research labs to mainstream business adoption, with hyperscale cloud providers and enterprises racing to deploy models in fields ranging from healthcare and finance to autonomous systems and industrial automation. However, training and running these models requires enormous compute power, leading to spiraling energy demand.
Estimates suggest that a single training run for a state-of-the-art AI model can consume megawatt-hours of electricity, while datacentres powering global AI workloads are already responsible for hundreds of terawatt-hours annually. This raises pressing concerns about sustainability, carbon emissions, and infrastructure scalability.
By combining green hydropower-fed datacentres with ultra-efficient optical processors, Terakraft and Neurophos are positioning themselves as pioneers in tackling this global challenge.
Terakraft: green datacentres with benchmark efficiency
Based in Norway, Terakraft operates renewably powered datacentres designed for minimal environmental impact. Its infrastructure is powered entirely by hydropower and cooled using natural lake water, enabling a power usage effectiveness (PUE) below 1.1—a figure that ranks among the most efficient in the global industry.
Key highlights of Terakraft’s approach include:
Hydropower utilization: 100% renewable energy supply, aligning with Europe’s carbon-neutral targets.
Natural cooling: Use of lake water eliminates the need for power-hungry cooling systems.
Circular construction: Repurposing an existing reinforced concrete hydropower plant, avoiding the embodied carbon associated with new builds.
By designing infrastructure that reuses and optimizes existing assets, Terakraft is setting a new benchmark for low-emission datacentre operations.
Neurophos: optical chips for 100x efficiency
On the hardware front, Neurophos is developing optical processing units (OPUs) designed to deliver orders-of-magnitude energy savings compared to traditional GPUs. Its proprietary technology is based on the miniaturization of optical modulators by 10,000x, enabling optical systems to scale down to chip-level integration.
Neurophos claims its OPUs can deliver:
100x efficiency gains over leading GPUs on equivalent AI inference workloads.
Performance density of 100 GPUs, consuming just 1% of the energy.
Compute-in-memory architecture inspired by the human brain, reducing latency and boosting scalability.
Ultra-fast inference validated through end-to-end simulation results.
This approach addresses one of AI’s biggest pain points: the gap between compute demand and sustainable energy supply. By embedding optical processing in memory, Neurophos mimics neural efficiency while enabling ultra-dense, low-power computation.
A pilot for real-world validation
As part of their collaboration, Terakraft and Neurophos will launch a commercial early access program in 2027, hosting Neurophos’ accelerated inference platform in Terakraft’s Norwegian datacentre.
This pilot will serve as a real-world proving ground for sustainable AI compute, giving select enterprise clients access to:
Ultra-efficient inference workloads running on Neurophos’ OPUs.
Carbon-neutral datacentre infrastructure powered by renewable hydropower.
Operational benchmarks to validate the scalability of green AI deployments.
The pilot is designed to test not only technical feasibility but also commercial viability, demonstrating how next-generation chips and renewable infrastructure can align to meet enterprise AI needs.
Leadership perspectives
Terakraft chairman Giorgio Sbriglia emphasized the alignment between sustainability and performance:
“By hosting Neurophos’ ultra-efficient optical chips in our green datacentre for select enterprise clients, we not only reduce our carbon footprint but also raise the bar for energy-efficient AI infrastructure. Our mission has always been to power the future responsibly, and this collaboration brings that vision to life.”
Neurophos CEO and founder Patrick Bowen highlighted the shared mission to democratize AI sustainably:
“Terakraft’s commitment to renewable energy and innovative technologies aligns perfectly with our mission to democratise high-performance AI. By deploying our 100x more efficient inference chips in Terakraft’s green datacentre we’re proving that AI’s exponential growth can be achieved sustainably, together.”
Terakraft and Neurophos carve a niche in sustainable AI
The collaboration between Terakraft and Neurophos places them at the intersection of two of the most pressing challenges in technology: scaling AI compute power while maintaining sustainability. Their combination of renewable-powered datacentres and optical AI chips differentiates them from the traditional ecosystem dominated by hyperscalers and GPU incumbents.
Hyperscaler giants: efficiency within limits
Global leaders in AI infrastructure—Microsoft (Azure), Google Cloud, Amazon Web Services, and Meta—have invested heavily in renewable energy purchasing agreements, energy-efficient datacentre designs, and AI-optimized silicon. However, they still rely primarily on NVIDIA’s GPU clusters and other conventional processors, which consume significant power despite efficiency gains.
Microsoft: Partnering with OpenAI, it has scaled Azure’s GPU superclusters while exploring liquid cooling and immersion cooling.
Google: Developing custom TPUs (tensor processing units) and promoting carbon-neutral datacentres but still constrained by energy use per operation.
Amazon: Expanding renewable-backed infrastructure and investing in its Inferentia and Trainium chips but facing challenges scaling efficiently across diverse workloads.
While hyperscalers have optimized operations, their reliance on electrical processing architectures limits the magnitude of efficiency gains. Terakraft and Neurophos offer a step-change alternative by tackling energy efficiency at the chip architecture level while embedding it within renewable-powered infrastructure.
Hardware innovators: the race beyond GPUs
Alongside NVIDIA, several hardware players are racing to design chips optimized for AI workloads:
NVIDIA: Dominates with GPUs like the H100, but faces scrutiny over its growing energy demands.
AMD: Offers competing AI accelerators with a focus on performance-per-watt improvements.
Cerebras: Pioneers wafer-scale engines for massive parallelism, pushing boundaries in training speed but still reliant on conventional energy-intensive silicon.
Graphcore and SambaNova: Developing AI-specialized architectures but struggling to break NVIDIA’s market dominance.
What sets Neurophos apart is its claim of 100x efficiency through optical processing, miniaturized modulators, and compute-in-memory designs. This represents a paradigm shift rather than an incremental improvement—potentially leapfrogging traditional GPU and accelerator roadmaps if scalability is proven.
Sustainable datacentres: Europe’s differentiator
Terakraft’s Norwegian base provides a natural strategic edge. With abundant hydropower and cold climates, Scandinavia has become a hub for green datacentre projects. Competitors in this space include:
Green Mountain (Norway): Hydropower and fjord-cooled datacentres.
EcoDataCenter (Sweden): Uses renewable energy and recycles waste heat.
Kolos (Norway): Marketed as one of the world’s largest renewable datacentres.
Terakraft’s advantage lies in its repurposing strategy—reusing an existing hydropower plant to avoid the embodied carbon associated with new construction. This adds another layer of sustainability beyond renewable energy sourcing.
Positioning against incumbents
The Terakraft–Neurophos model contrasts sharply with today’s hyperscaler-driven AI build-outs:
Traditional path: Incremental improvements in cooling, power sourcing, and chip design.
Their model: Fundamental architectural shift in compute (optical processors) plus end-to-end green infrastructure.
This positions them not as direct competitors to the hyperscalers but as disruptive enablers—potentially supplying sustainable compute platforms for enterprises seeking alternatives to carbon-heavy AI clouds.
Strategic opportunities
The collaboration could attract:
European enterprises: Firms under regulatory and ESG pressure may prefer locally powered, carbon-efficient AI infrastructure.
Governments: Seeking sovereign, sustainable compute capacity aligned with EU’s Green Deal and digital sovereignty goals.
Hyperscalers themselves: NVIDIA and Microsoft have already invested in startups like Ayar Labs (optical I/O). If Neurophos proves its efficiency claims, hyperscalers may adopt rather than compete.
Challenges ahead
Despite strong positioning, Terakraft and Neurophos face risks:
Scalability: Neurophos must prove its optical chips can scale beyond lab demonstrations.
Capital intensity: Building commercial-scale datacentres and chip fabs requires billions, far beyond early pilots.
Market inertia: Enterprises and hyperscalers are deeply entrenched with NVIDIA GPUs, making displacement difficult.
Timeline: With the first pilot only slated for 2027, competitors may accelerate their own green AI solutions in the interim.
Strategic outlook
By uniting green infrastructure with ultra-efficient chips, Terakraft and Neurophos present a visionary model for sustainable AI. If successful, they could redefine how datacentres are built and powered—shifting the debate from incremental energy reductions to radical efficiency breakthroughs.
Rather than challenging hyperscalers head-on, they may carve out a sustainable premium niche, providing compute for enterprises, governments, and industries that demand both AI performance and carbon accountability.
Powered by Froala Editor