Tokens Per Watt Over Peak Performance: Retrofitting Heterogeneous Silicon for Production Inference
π This is Part 2 of an editorial series on the evolving economics of AI inference. If you missed Part 1, where I broke down the shift toward full-system integration, the lessons from the AI Infra Summit, and the βno forksβ open-source philosophy, you can read it HERE: Beyond the Accelerator: Why Silicon Challengers Must Transition to Full-System Infrastructure.
π μ΄κ²μ AI μΆλ‘ (Inference)μ μ§ννλ κ²½μ νμ κ΄ν μ¬μ€ μ리μ¦μ μ 2λΆμ λλ€. μ 체 μμ€ν ν΅ν©μ ν₯ν μ ν, AI μΈνλΌ μλ°(AI Infra Summit)μμμ κ΅ν, κ·Έλ¦¬κ³ "λ Έ ν¬ν¬(no forks)" μ€νμμ€ μ² νμ λ€λ£¬ μ 1λΆλ₯Ό λμΉμ ¨λ€λ©΄ μ¬κΈ°μμ μ½μ΄λ³΄μ€ μ μμ΅λλ€: κ°μκΈ°λ₯Ό λμ΄: μ€λ¦¬μ½ λμ μλ€μ΄ ν μμ€ν μΈνλΌλ‘ μ νν΄μΌ νλ μ΄μ .
In my previous post, we explored why the next phase of the AI infrastructure market will be dominated by complete, integrated systems rather than standalone accelerators. But as enterprises transition from casual AI experimentation to large-scale production, the underlying technical arguments must answer to the realities of the corporate balance sheet. This post continues on with the Rebellions story.
μ΄μ κΈμμλ AI μΈνλΌ μμ₯μ λ€μ λ¨κ³κ° λ 립ν κ°μ기보λ€λ μμ νκ³ ν΅ν©λ μμ€ν μ μν΄ μ£Όλλ μ΄μ λ₯Ό μ΄ν΄λ³΄μμ΅λλ€. κ·Έλ¬λ κΈ°μ λ€μ΄ κ°λ²Όμ΄ AI μ€ν λ¨κ³μμ λκ·λͺ¨ νλ‘λμ (μμ© μ΄μ) λ¨κ³λ‘ μ νν¨μ λ°λΌ, κ·Έ μ΄λ©΄μ κΈ°μ μ μ£Όμ₯μ κΈ°μ μ μ¬λ¬΄μ νκ° μ§λ©΄ν νμ€μ λΆν©ν΄μΌ ν©λλ€. μ΄λ² κΈμ 리벨리μ¨(Rebellions)μ μ΄μΌκΈ°λ‘ κ³μλ©λλ€.
The mathematical argument for specialized inference hardware is becoming difficult for CFOs to ignore. Recent platform data reveals aggressive competitive claims: delivering roughly three times more tokens per dollar, more than double the tokens per watt, and a threefold reduction in overall cost per rack compared to legacy NVIDIA GPU configurations. Just as compelling is the physical profile of the system, which logs a highly optimized four kilowatts of power consumption per deployment unit.
νΉνλ μΆλ‘ νλμ¨μ΄λ₯Ό λ·λ°μΉ¨νλ μνμ λ Όκ±°λ μ΅κ³ μ¬λ¬΄μ± μμ(CFO)λ€μ΄ 무μνκΈ° μ΄λ €μ΄ μμ€μ΄ λκ³ μμ΅λλ€. μ΅κ·Ό νλ«νΌ λ°μ΄ν°λ κΈ°μ‘΄ μλΉλμ(NVIDIA) GPU ꡬμ±μ λΉν΄ λ¬λ¬λΉ μ²λ¦¬ ν ν° μκ° μ½ 3λ°° λ§κ³ , μνΈλΉ ν ν° μλ 2λ°° μ΄μμ΄λ©°, λλΉ μ 체 λΉμ©μ 3λΆμ 1λ‘ μ κ°λλ€λ 곡격μ μΈ κ²½μλ ₯ μ§νλ₯Ό 보μ¬μ€λλ€. λ°°ν¬ μ λλΉ μ λ ₯ μλΉλμ΄ 4ν¬λ‘μνΈλ‘ λ§€μ° μ΅μ νλμ΄ μλ€λ μμ€ν μ 물리μ νλ‘ν λν λ§μ°¬κ°μ§λ‘ λ§€λ ₯μ μ λλ€.
While these metrics originate as competitive claims, their broader strategic implication is what matters to infrastructure architects. In the macro-scale game, operating costs dictate survival. When your infrastructure scales across thousands of nodes, the cascading costs of thermal management, floor space, and raw utility draws completely dictate the viability of your AI business model. Peak lab performance becomes a secondary luxury if the baseline economics of a deployment break your budget.
μ΄λ¬ν μ§νλ€μ΄ κ²½μμ¬ μ£Όμ₯μμ λΉλ‘―λ κ²μ΄λΌ νλλΌλ, μΈνλΌ μν€ν νΈλ€μκ² μ€μν κ²μ κ·Έ μ΄λ©΄μ λ΄κΈ΄ κ΄λ²μν μ λ΅μ μλ―Έμ λλ€. κ±°μμ κ·λͺ¨μ κ²½μμμ μ΄μ λΉμ©μ μμ‘΄μ μ’μ°ν©λλ€. μΈνλΌκ° μμ² κ°μ λ Έλλ‘ νμ₯λ λ, λ°μ΄ κ΄λ¦¬, λ°λ₯ λ©΄μ , μ λ ₯ μλΉλμμ λ°μνλ μ°μμ λΉμ©μ AI λΉμ¦λμ€ λͺ¨λΈμ 쑴립 μ¬λΆλ₯Ό μμ ν κ²°μ μ§μ΅λλ€. λ°°ν¬μ κΈ°λ³Έ κ²½μ μ±μ΄ μμ°μ μ΄κ³Όν΄ λ²λ¦°λ€λ©΄, μ΅κ³ μμ€μ μ€νμ€ μ±λ₯μ λΆμ°¨μ μΈ μ¬μΉμ λΆκ³Όν©λλ€.
Why Benchmarks Distort Production Truths
One of the most vital insights regarding current market positioning is a healthy skepticism toward standard synthetic benchmarking. This perspective is born from deep institutional experience; leadership teams managing these hardware cycles have spent decades evaluating silicon.
νμ¬ μμ₯ ν¬μ§μ λκ³Ό κ΄λ ¨ν΄ κ°μ₯ μ€μν ν΅μ°° μ€ νλλ νμ€μ μΈ ν©μ± λ²€μΉλ§ν¬(Synthetic benchmarking)μ λν 건μ ν νμλ‘ μ λλ€. μ΄λ¬ν μκ°μ νλμ¨μ΄ μ£ΌκΈ°λ₯Ό μμ λ κ° νκ°ν΄ μ¨ λ¦¬λμ νλ€μ κΉμ νμ₯ κ²½νμμ λΉλ‘―λμμ΅λλ€.
Synthetic benchmarks are highly controlled laboratory experiments. They measure isolated performance under variables such as specific model weights, predictable batch sequences, and pristine datasets. Production environments, by contrast, are messy, unpredictable, and highly volatile. Real-world enterprise deployments require flawless uptime, predictable scaling curves under erratic user spikes, robust security profiles, and enterprise-grade resilience. Most importantly, they have to support a viable business proposition.
ν©μ± λ²€μΉλ§ν¬λ κ³ λλ‘ ν΅μ λ μ€νμ€ μ€νμ λλ€. μ΄λ νΉμ λͺ¨λΈ κ°μ€μΉ, μμΈ‘ κ°λ₯ν λ°°μΉ μνμ€, κ·Έλ¦¬κ³ μ μ λ λ°μ΄ν°μ κ°μ λ³μλ€ μμμ κ³ λ¦½λ μ±λ₯μ μΈ‘μ ν©λλ€. λ°λ©΄ νλ‘λμ νκ²½μ κ·Έμ λ¬λ¦¬ 볡μ‘νκ³ μμΈ‘ λΆκ°λ₯νλ©° λ³λμ±μ΄ ν½λλ€. μ€μ κΈ°μ λ°°ν¬ νκ²½μμλ μλ²½ν κ°λ μκ°, μμΈ‘ λΆκ°λ₯ν μ¬μ©μ κΈμ¦ μν©μμλ νλ€λ¦¬μ§ μλ μμΈ‘ κ°λ₯ν νμ₯ 곑μ , κ°λ ₯ν 보μ νλ‘ν, κ·Έλ¦¬κ³ κΈ°μ μ© μμ€μ νλ ₯μ±μ΄ μꡬλ©λλ€. 무μ보λ€λ, μ€ν κ°λ₯ν λΉμ¦λμ€ λͺ λΆμ λ·λ°μΉ¨ν΄μΌ ν©λλ€.
The pitfalls of chasing superficial metrics can be seen in a stark production anecdote: a development organization observed an order-of-magnitude spike in the volume of raw code generated by its AI tools. However, when management looked closer at the actual production environment, the volume of code committed and deployed remained completely flat.
νλ©΄μ μΈ μ§νλ₯Ό μ’μ λμ ν¨μ μ ν κ°λ° μ‘°μ§μ κ·Ήλ¨μ μΈ νλ‘λμ μΌνμμ μ λλ¬λ©λλ€. AI λꡬμ μν΄ μμ±λ μμ μ½λμ μμ΄ μλ¦Ώμ λ¨μλ‘ κΈμ¦ν κ²μ λͺ©κ²©ν κ²μ λλ€. κ·Έλ¬λ κ²½μμ§μ΄ μ€μ νλ‘λμ νκ²½μ μμΈν λ€μ¬λ€λ³΄μμ λ, μ€μ λ‘ μ»€λ°λκ³ λ°°ν¬λ μ½λμ μμ μμ ν μ 체λμ΄ μμμ΅λλ€.
The lesson is a vital reality check for the C-suite: churning out more tokens does not inherently equal true enterprise productivity. For organizations scaling AI across their operations, the metric that matters is business outcome. In other words, what the infrastructure actually empowers the enterprise to achieve.
μ΄ κ΅νμ C-suite(μ΅κ³ κ²½μμ§)μκ² κΌ νμν νμ€μ κ²½κ°μ¬μ μ€λλ€. λ¨μν λ λ§μ ν ν°μ μ°μ΄λ΄λ κ²μ΄ κ³§ μ§μ ν κΈ°μ μμ°μ±μΌλ‘ μ§κ²°λλ κ²μ μλλλ€. μ€νΌλ μ΄μ μ λ°μ κ±Έμ³ AIλ₯Ό νμ₯νλ μ‘°μ§μκ² μ€μν μ§νλ λΉμ¦λμ€ κ²°κ³Όλ¬Όμ λλ€. μ¦, μΈνλΌκ° μ€μ λ‘ κΈ°μ μ μ±μ·¨λ₯Ό μ΄λ»κ² λ·λ°μΉ¨νλλ νλ κ²μ λλ€.
Production Proof: From SK Telecom to Sovereign Air-Gaps
Rather than building complex infrastructure in a vacuum and searching for a problem to solve, Rebellions and SK Telecom built their framework around a distinct internal corporate bottleneck: customer service applications and heavy textual workloads.
볡μ‘ν μΈνλΌλ₯Ό μ§κ³΅ μνμμ ꡬμΆν λ€ ν΄κ²°ν΄μΌ ν λ¬Έμ λ₯Ό μ°Ύμ λμλ λμ , 리벨리μ¨κ³Ό SKν λ μ½€μ λͺ νν μ¬λ΄ λΉμ¦λμ€ λ³λͺ© νμ, μ¦ κ³ κ° μΌν° μ ν리μΌμ΄μ κ³Ό λ¬΄κ±°μ΄ ν μ€νΈ μν¬λ‘λλ₯Ό μ€μ¬μΌλ‘ νλ μμν¬λ₯Ό ꡬμΆνμ΅λλ€.
The resulting deployment bypassed the standard GPU route, creating an operational profile that proved faster, cheaper, and vastly more scalable than the telecom giantβs previous infrastructure. Today, that foundational architecture has evolved into a full-scale external commercial service, processing roughly 50 million API calls per day for businesses across South Korea.
κ·Έ κ²°κ³Ό νμν λ°°ν¬ νκ²½μ νμ€μ μΈ GPU κ²½λ‘λ₯Ό μ°ννμ¬, μ΄ ν΅μ κ±°μΈμ μ΄μ μΈνλΌλ³΄λ€ λ λΉ λ₯΄κ³ μ λ ΄νλ©° ν¨μ¬ νμ₯μ±μ΄ λ°μ΄λ μ΄μ νλ‘νμ μ μ¦ν΄ λμ΅λλ€. μ€λλ κ·Έ κΈ°μ΄ μν€ν μ²λ μ λ©΄μ μΈ λμΈ μμ© μλΉμ€λ‘ μ§ννμ¬, νκ΅ μ μμ κΈ°μ λ€μ μν΄ ν루μ μ½ 5,000λ§ κ±΄μ API νΈμΆμ μ²λ¦¬νκ³ μμ΅λλ€.
This same playbook is driving their push into Sovereign AIβgiving local organizations and governments absolute custody over their computing stacks. In partnership with KT Cloud, Rebellions has designed a secure, air-gapped AI infrastructure solution specifically tailored for mission-critical government workloads. Built entirely at rack scale and utilizing OpenStack for robust enterprise orchestration, the system is designed for environments where data security and operational isolation are non-negotiable. By building systems capable of surviving the strict parameters of national government workloads, a highly reliable blueprint is established for enterprise clients in highly regulated sectors like banking, healthcare, and critical infrastructure.
μ΄μ λμΌν νλ μ΄λΆμ΄ λ‘컬 μ‘°μ§κ³Ό μ λΆκ° μ»΄ν¨ν μ€νμ λν μ λμ μΈ ν΅μ κΆμ κ°λλ‘ νλ 'μλ²λ¦° AI(Sovereign AI)' λΆμΌλ‘μ μ§μΆμ μ΄λκ³ μμ΅λλ€. KTν΄λΌμ°λμμ ννΈλμμ ν΅ν΄ 리벨리μ¨μ μ λΆμ λ―Έμ ν¬λ¦¬ν°μ»¬ μν¬λ‘λμ νΉνλ μμ νκ³ μμ΄κ°(μΈλΆλ§ μ°¨λ¨)λ AI μΈνλΌ μ루μ μ μ€κ³νμ΅λλ€. μμ ν λ λ¨μλ‘ κ΅¬μΆλμμΌλ©° κ°λ ₯ν κΈ°μ μ© μ€μΌμ€νΈλ μ΄μ μ μν΄ μ€νμ€ν(OpenStack)μ νμ©ν μ΄ μμ€ν μ λ°μ΄ν° 보μκ³Ό μ΄μ κ²©λ¦¬κ° μ λμ μΌλ‘ ννλ μ μλ νκ²½μ μν΄ μ€κ³λμμ΅λλ€. κ΅κ° μ λΆ κΈ°κ΄μ μ격ν λ§€κ°λ³μλ₯Ό 견λλΌ μ μλ μμ€ν μ ꡬμΆν¨μΌλ‘μ¨, λ± νΉ, ν¬μ€μΌμ΄, ν΅μ¬ μΈνλΌμ κ°μ΄ κ·μ κ° μ격ν λΆμΌμ κΈ°μ κ³ κ°λ€μ μν λμ μ λ’°μ±μ μ²μ¬μ§μ΄ λ§λ ¨λ κ²μ λλ€.
The Case for Heterogeneous AI
Underlying this entire strategy is a broader, inevitable prediction about the future of AI infrastructure: computing will become increasingly heterogeneous. The idea that a single type of monolithic processor should dominate every single layer of the AI computing stack is an antiquated relic of early market architecture. Different workloads, fluctuating sequence lengths, and varied application budgets require a flexible blend of processing units, systems, models, and software.
μ΄ μ 체 μ λ΅μ κΈ°μ μλ AI μΈνλΌμ λ―Έλμ λν λ³΄λ€ κ΄λ²μνκ³ νΌν μ μλ μμΈ‘μ΄ κΉλ € μμ΅λλ€. λ°λ‘ μ»΄ν¨ν μ΄ μ μ λ μ΄μ’ ν(Heterogeneous)λ κ²μ΄λΌλ μ μ λλ€. λ¨μΌ μ νμ λ¨μΌμ²΄(Monolithic) νλ‘μΈμκ° AI μ»΄ν¨ν μ€νμ λͺ¨λ λ¨κ³λ₯Ό μ§λ°°ν΄μΌ νλ€λ μκ°μ μ΄κΈ° μμ₯ μν€ν μ²μ ꡬμλμ μ λ¬Όμ λλ€. μλ‘ λ€λ₯Έ μν¬λ‘λ, λ³λνλ μνμ€ κΈΈμ΄, κ·Έλ¦¬κ³ λ€μν μ ν리μΌμ΄μ μμ°μλ νλ‘μΈμ± μ λ, μμ€ν , λͺ¨λΈ, κ·Έλ¦¬κ³ μννΈμ¨μ΄μ μ μ°ν μ‘°ν©μ΄ νμν©λλ€.
This philosophy is steering international footprints, highlighted by Rebellionsβ partnership with AIand, a prominent Japan-based AI inference service provider focused entirely on heterogeneous infrastructure. The ambitious rollout plans call for deployment across five independent data centers, introducing 40 megawatts of raw capacity and up to 100 integrated rack-scale systems.
μ΄λ¬ν μ² νμ μ΄μ’ μΈνλΌμ μ μ μΌλ‘ μ΄μ μ λ§μΆ μΌλ³Έ κΈ°λ°μ μ λ ₯ AI μΆλ‘ μλΉμ€ μ 곡μ¬μΈ AIandμμ ννΈλμμ ν΅ν΄ κ°μ‘°λλ―μ΄, κ΅μ μ μ μ§λ₯Ό μ£Όλνκ³ μμ΅λλ€. μΌμ¬ μ°¬ μΆμ κ³νμ λ°λΌ 5κ°μ λ 립λ λ°μ΄ν° μΌν°μ κ±Έμ³ 40λ©κ°μνΈμ μμ μ©λκ³Ό μ΅λ 100κ°μ ν΅ν© λ μ€μΌμΌ μμ€ν μ λ°°ν¬ν μμ μ λλ€.
In this heterogeneous future, the power shifts back to the operators. Infrastructure teams will no longer be locked into a single ecosystem; instead, they can select the most appropriate accelerator for a particular application based on performance, economics, power consumption, availability, and operational requirements. Crucially, such an approach allows organizations to make greater use of their existing legacy infrastructure while adding new capabilities incrementally, rather than forcing a costly rip-and-replace architecture.
μ΄λ¬ν μ΄μ’ νλ λ―Έλμμλ μ£ΌλκΆμ΄ λ€μ μ΄μμμκ² λμ΄κ°λλ€. μΈνλΌ νλ€μ λ μ΄μ λ¨μΌ μνκ³μ κ°ν μ§λ΄μ§ μκ³ , μ±λ₯, κ²½μ μ±, μ λ ₯ μλΉ, κ°μ©μ± λ° μ΄μ μꡬμ¬νμ λ°λΌ νΉμ μ ν리μΌμ΄μ μ κ°μ₯ μ ν©ν κ°μκΈ°λ₯Ό μ νν μ μκ² λ©λλ€. κ²°μ μ μΌλ‘ μ΄λ¬ν μ κ·Ό λ°©μμ ν΅ν΄ μ‘°μ§λ€μ λΉμ©μ΄ λ§μ΄ λλ μ λ©΄ κ΅μ²΄(Rip-and-replace) μν€ν μ²λ₯Ό κ°μ λΉνλ λμ , μλ‘μ΄ κΈ°λ₯μ μ μ§μ μΌλ‘ μΆκ°νλ©΄μ κΈ°μ‘΄μ λ κ±°μ μΈνλΌλ₯Ό λ ν¨κ³Όμ μΌλ‘ νμ©ν μ μμ΅λλ€.
Designing for Deployment, Not Experiments
Ultimately, the market is moving past the phase of unconstrained AI experimentation and stepping firmly into large-scale production. Performance will always remain a pillar of the equation. But as inference workloads scale, organizations must evaluate the complete picture: power constraints, capital expenditure, operating expenditure, physical space, reliability, software compatibility, and business outcomes.
κΆκ·Ήμ μΌλ‘ μμ₯μ μ μ½ μλ AI μ€ν λ¨κ³λ₯Ό μ§λ λκ·λͺ¨ νλ‘λμ λ¨κ³λ‘ λ¨λ¨ν λ°μ λ€μ¬λκ³ μμ΅λλ€. μ±λ₯μ μΈμ λ λ°©μ μμ ν΅μ¬ κΈ°λ₯μΌλ‘ λ¨μ κ²μ λλ€. κ·Έλ¬λ μΆλ‘ μν¬λ‘λκ° νμ₯λ¨μ λ°λΌ μ‘°μ§λ€μ μ λ ₯ μ μ½, μλ³Έμ μ§μΆ(CapEx), μ΄μμ μ§μΆ(OpEx), 물리μ 곡κ°, μ λ’°μ±, μννΈμ¨μ΄ νΈνμ±, κ·Έλ¦¬κ³ λΉμ¦λμ€ μ±κ³ΌλΌλ μ 체 κ·Έλ¦Όμ νκ°ν΄μΌ ν©λλ€.
The pitch is less about chasing a fleeting headline benchmark and far more about the unglamorous work of deployment. AI inference has officially become an infrastructure category in its own right. The winners in this next phase of the market will have to compete on economics, raw efficiency, openness, and the uncompromising ability to deliver real-world business value.
μ΄λ² νΌμΉλ μΌμμ μΈ ν€λλΌμΈ λ²€μΉλ§ν¬λ₯Ό μ«λ κ²μ΄λΌκΈ°λ³΄λ€λ λ°°ν¬λΌλ λΉλμ§ μμ§λ§ 묡묡ν μμ μ ν¨μ¬ κ°κΉμ΅λλ€. AI μΆλ‘ μ 곡μμ μΌλ‘ λ μμ μΈ μΈνλΌ μΉ΄ν κ³ λ¦¬κ° λμμ΅λλ€. μ΄ μμ₯μ λ€μ λ¨κ³μμ μΉλ¦¬νλ μ΄λ€μ κ²½μ μ±, μμ ν¨μ¨μ±, κ°λ°©μ±, κ·Έλ¦¬κ³ μ€μΈκ³μ λΉμ¦λμ€ κ°μΉλ₯Ό μ λ¬ν μ μλ νν μλ μλμ λ°νμΌλ‘ κ²½μν΄μΌ ν κ²μ λλ€.
Visit Www.Rebellions.ai to learn more.
μμΈν λ΄μ©μ Www.Rebellions.aiλ₯Ό λ°©λ¬Έν΄ νμΈν΄ 보μΈμ.