JarvisLabs Managed Endpoints: E2E Networks’ Bid to Simplify Open-Weight AI Deployment

JarvisLabs Managed Endpoints launched on September 7, 2026, as a service that packages model setup, validation, and inference optimization into ready-to-deploy API endpoints for open-weight AI models. The platform, owned by E2E Networks Limited, targets enterprises that spend weeks configuring GPU infrastructure before a newly released model can handle production workloads.

The JarvisLabs Managed Endpoints launch arrives at a moment when open-weight model adoption is accelerating but deployment complexity remains a bottleneck. Enterprises want to run models like DeepSeek V4 Flash and GLM-5.3 without hiring specialized inference engineers — and JarvisLabs is betting that pre-validated, pre-tuned endpoints can eliminate the 15 to 20 working days that typically separate a model release from production readiness.

Layer 1 — Why Now: The Open-Weight Inference Boom

The Deployment Bottleneck

The press release identifies a specific pain point: getting a newly released AI model ready for production takes 15 to 20 working days, even with JarvisLabs‘ own AI research team working alongside customers. Teams must evaluate GPUs, serving software, and model-specific settings before an endpoint can reliably handle production traffic.

This bottleneck is real and widely recognized. The serverless inference market for open-weight models has consolidated around seven major providers by Q2 2026 — Together AI, Fireworks AI, Groq, Cerebras, Replicate, OctoAI, and Anyscale — according to Digital Applied. Pricing on the same model spreads 6x across providers, and latency spreads 5 to 7x. The market is mature enough that the choice is no longer whether to use managed inference, but which provider and deployment model fits best.

E2E Networks’ Growth Trajectory

E2E Networks, JarvisLabs’ parent company, is experiencing explosive growth. Q1 FY27 revenue reached Rs 156.8 crore, up 334% year-on-year and 64% quarter-on-quarter. EBITDA margin expanded to 75.2%, with Q1 EBITDA alone reaching 93% of full-year FY26 EBITDA. The company swung to a net profit of Rs 43.9 crore in Q1 FY27, compared to a net loss of Rs 2.8 crore in the year-ago quarter.

This growth is driven by the B200 cluster going live on the TIR platform and GPU infrastructure scaling to approximately 5,100 GPUs. E2E Networks raised over Rs 15,919 crore through preferential allotments, including a strategic investment of Rs 1,079 crore from Larsen and Toubro. The company positions itself as India’s sovereign AI infrastructure provider, with nearly 3,000 Hopper GPUs deployed and NVIDIA Blackwell installations scaling.

India’s AI Infrastructure Push

India’s data centre pipeline stands at 8.33 GW, more than five times the approximately 1.6 GW of live capacity today, according to Knight Frank India. Around $30 billion of investment underpins this expansion. The IndiaAI Mission has over 38,000 GPUs. Oppenheimer Research estimates the global sovereign AI infrastructure opportunity at $1.5 trillion this decade, while McKinsey estimates $6.7 trillion in global data centre capex will be required by 2030, with around 70% driven by AI.

E2E Networks was awarded a contract under the IndiaAI Mission and powers foundational models for the program. The Managed Endpoints launch extends this sovereign positioning into the enterprise inference market.

How JarvisLabs Managed Endpoints Compare to Together AI, Fireworks AI, and Replicate

Together AI: The Catalog Leader

Together AI offers serverless inference across 200 to 500+ open-source models, covering Llama, DeepSeek, GLM, Qwen, and GPT-OSS families. Pricing starts at $0.10 per million tokens for smaller models, with mid-range models like Llama 70B at approximately $0.88 per million tokens. Together AI supports LoRA and full fine-tuning with downloadable checkpoints, dedicated H100 endpoints starting at $2.99 per hour, and batch processing at 50% off. The platform provides an OpenAI-compatible API, meaning existing code works with minimal changes.

Together AI raised significant venture capital and built a research team focused on inference optimization. The company claims up to 2.75x faster serverless inference compared to competitors.

Fireworks AI: The Speed Leader

Fireworks AI, built by former Meta PyTorch engineers, differentiates on inference speed and serving-path control. Its FireAttention serving stack delivers the lowest P99 latency on open-source models, often 2 to 5x faster than competitors. Fireworks offers Standard, Priority, and Fast serving tiers, documented zero-data-retention defaults, and a US-only serverless option for compliance-sensitive workloads.

Pricing uses size-tiered serverless rates — models grouped into parameter bands with every model in a band costing the same. LoRA fine-tuning is supported, and serving a fine-tuned LoRA costs the same per token as the base model. GPT-OSS 120B lists at $0.15 per million input tokens and $0.60 per million output tokens.

Groq and Cerebras: The Hardware Specialists

Groq pushes a differentiated hardware story around its LPU system, delivering 500+ tokens per second on 70B-class models. Cerebras uses wafer-scale chips for similar throughput. Both charge more per token but deliver 5 to 7x throughput compared to commodity H100 endpoints. Neither supports fine-tuning on standard accounts.

Replicate: The Breadth Play

Replicate hosts 2,000+ models covering LLMs, image generation, video, audio, and speech. Its per-second billing model suits intermittent workloads, and custom models can be deployed using the Cog packaging tool. Cold starts remain a challenge — models not kept hot can take 30+ seconds to wake up.

Where JarvisLabs Fits

JarvisLabs occupies a different position from all these providers. It is not a US-based serverless inference platform. As an Indian GPU cloud, now backed by E2E Networks’ 5,100+ GPU fleet, it offers managed endpoints with a specific value proposition: sovereignty, data location control, and pricing in rupees.

The platform rents NVIDIA GPUs at rates starting from Rs 38.88 per hour, 60 to 70% cheaper than hyperscalers according to founder Vishnu Subramanian. Managed Endpoints extends this cost advantage into the managed inference space — but with a model catalog of five models, compared to Together AI’s 200+ or Replicate’s 2,000+.

The competitive question is whether enterprises will choose a sovereign Indian provider with a smaller catalog but localized control over a US-based platform with broader model coverage and deeper optimization research.

Layer 3 — Public-Data Sweep: E2E Networks and JarvisLabs by the Numbers

E2E Networks Financial Profile

E2E Networks Limited trades on the NSE (symbol: E2E) and BSE (code: 544783). Founded in 2009 by Tarun Dua, the company went public on NSE Emerge in 2018 at an IPO price of Rs 57 per share, oversubscribed 70x. It migrated to the NSE mainboard in 2022 and listed on BSE mainboard in June 2026.

Financial trajectory: FY23 revenue of Rs 66.2 crore grew to Rs 94.5 crore in FY24, Rs 164 crore in FY25 (+74% YoY), and Rs 245.6 crore in FY26 (+50% YoY). FY26 saw a net loss of Rs 15.6 crore due to depreciation surging 182% to Rs 169.2 crore on new GPU capex, despite EBITDA growing 31% to Rs 126.3 crore. Q1 FY27 marked a dramatic turnaround: revenue of Rs 156.8 crore, EBITDA of Rs 117.9 crore at 75.2% margin, and PAT of Rs 43.9 crore.

The company has 213 employees, 50% women on the board, and zero regulatory penalties. Monthly recurring revenue reached Rs 11.2 crore as of March 2025, up from Rs 10.9 crore in March 2024.

JarvisLabs Background

JarvisLabs was founded in 2019 by Vishnu Subramanian in Coimbatore, on the outskirts of the city following what he calls the “Zoho School of Thought.” The company started with four people and one GPU bought for Rs 1.5 lakh. Subramanian described the early days: “One of us literally sat inside it while we figured out how to assemble it.”

The platform went live in January 2021 as a one-click GPU cloud for AI researchers. It became part of Fast.ai’s suggested cloud providers list and adopted by 1,000+ customers globally. JarvisLabs was sold to E2E Networks in August 2025, and Subramanian now serves as Head of Product and Marketing at E2E Cloud while continuing as JarvisLabs founder.

The team has grown to approximately 10 people. Recent product launches include GPU VMs, serverless computing, Infiniband clusters, VPC and security groups, and the AI-native CLI — all shipped in the months leading up to the Managed Endpoints announcement.

Vishnu Subramanian’s Background

Subramanian is a Fast.ai alumnus and active Kaggle practitioner with 93 public GitHub repositories. Before JarvisLabs, he served as CTO at SmartNomad and Principal Solutions Architect at Affine Analytics. He has 15 years and 9 months of total professional experience. His GitHub repositories include “Deep learning with PyTorch Book” from Packt (91 stars) and a 21st-place solution for the TGS Salt Identification Challenge on Kaggle (85 stars).

On his personal blog, Subramanian wrote: “I was naive to start Jarvis Labs. If I had known how hard it would be to build a GPU platform from India, I probably would not have started. But that naivety also helped us begin.”

The Model Catalog

The Managed Endpoints catalog currently includes five models: DeepSeek V4 Flash, Gemma 4 31B, GLM-5.2, GLM-5.3, and GLM-5.3 Flash. These cover reasoning, multimodal, and tool-calling workloads. JarvisLabs plans to add speech-to-text, text-to-speech, and text-to-video models.

The validation behind each endpoint can be substantial. Preparing GLM-5.2 required more than 1,000 GPU-hours and $5,000 in compute on an eight-H200 node. The deployment scored 80.9% on Terminal-Bench 2.1, closely matching the 81.0% score reported by model developer Z.ai. This level of validation rigor is a genuine differentiator — most managed inference providers do not publicly disclose benchmark replication results.

E2E Networks’ Two AI Platforms

E2E Networks operates two AI platforms. TIR handles training, fine-tuning, retrieval-augmented generation, and model endpoints — positioned for enterprises and researchers. JarvisLabs serves as the developer-first GPU cloud, now extending into managed endpoints. The B200 cluster went live on the TIR platform in Q1 FY27, contributing to revenue within its first quarter.

Layer 4 — The Unasked Question: What the Press Release Avoids

Pricing for Managed Endpoints

The press release does not disclose pricing for Managed Endpoints. Competitor pricing is well-documented: Together AI charges $0.10 to $2.50 per million tokens, Fireworks AI charges $0.15 per million input tokens for GPT-OSS 120B, Groq charges $0.05 per million input tokens for Llama 3.1 8B. JarvisLabs’ raw GPU rental starts at Rs 38.88 per hour, but the managed endpoint pricing model — per-token, per-hour, or subscription — is not specified.

Model Catalog Breadth

Five models is a narrow catalog. Together AI offers 200 to 500+ models, Fireworks AI offers 100+, and Replicate offers 2,000+.

Even by the most conservative comparison, JarvisLabs has 40 to 100x fewer models than leading competitors. The press release states that new models will be added as the research team evaluates them, but no timeline or roadmap is provided.

Latency and Throughput Benchmarks

The press release shares one benchmark — 80.9% on Terminal-Bench 2.1 for GLM-5.2. But it does not provide latency metrics, tokens-per-second throughput, or cold-start times.

Fireworks AI publishes latency claims. Groq publishes 500+ tokens per second. Together AI claims 2.75x faster inference. JarvisLabs does not publish comparable performance data.

SLA and Availability

No service-level agreement, uptime guarantee, or availability commitment is mentioned. Together AI offers optional 99.9% uptime SLA on dedicated endpoints. Fireworks AI offers Standard, Priority, and Fast tiers. JarvisLabs’ Managed Endpoints do not specify any SLA structure.

Customer Testimonials and Case Studies

The press release includes no customer testimonials, case studies, or named enterprises using Managed Endpoints. The platform targets organizations running AI coding assistants, autonomous agents, enterprise copilots, and internal AI applications — but no specific customers are identified.

Data Sovereignty Claims

The press release mentions “control over data location” as a benefit but does not detail specific data residency guarantees, compliance certifications, or regulatory frameworks. E2E Networks positions as a sovereign cloud provider, but the Managed Endpoints announcement does not explicitly connect to this positioning.

Layer 5 — Honest Translation: What the Claims Mean

“Cut Weeks from Enterprise AI Deployment”

The claim that enterprises spend 15 to 20 working days getting a model production-ready is specific and credible. Subramanian says JarvisLabs observed this pattern “across enterprise deployments over the past several months.” The Managed Endpoints service performs this work once at the platform level, so customers receive a pre-validated endpoint.

This is a genuine value proposition. The question is whether it is unique. Together AI, Fireworks AI, and Replicate all offer pre-loaded models with no infrastructure management. The difference is that JarvisLabs publicly discloses the validation effort — 1,000+ GPU-hours and $5,000 for GLM-5.2 — while competitors do not.

“Ready-to-Deploy Endpoints for Leading Open-Weight AI Models”

The catalog includes DeepSeek V4 Flash, Gemma 4 31B, GLM-5.2, GLM-5.3, and GLM-5.3 Flash. These are legitimate open-weight models, but “leading” is a stretch when the catalog omits Llama, Qwen, Mistral, and GPT-OSS — all of which appear on competitor platforms. The catalog is focused rather than broad.

“Particularly Suited to Sustained Workloads”

This positioning distinguishes JarvisLabs from serverless providers like Replicate, which optimize for intermittent traffic. Sustained workloads with dedicated capacity are where E2E Networks’ GPU fleet creates a structural advantage — the company owns the GPUs and can allocate them to specific customers.

“Control Over Data Location Matters”

This is the sovereignty play. E2E Networks operates data centres in Noida and Chennai. For Indian enterprises concerned about data residency under the Digital Personal Data Protection Act, or for government agencies requiring domestic compute, this matters. But the press release does not make this connection explicit.

Layer 6 — Decision-Maker Framing: Who Should Care

For Enterprise AI Teams

If your team is deploying open-weight models and spending weeks on infrastructure configuration, Managed Endpoints offers a legitimate shortcut. The validation rigor — 1,000+ GPU-hours per model — suggests that the endpoints will work reliably. But you should evaluate whether the five-model catalog covers your use case, and whether the absence of published latency benchmarks meets your performance requirements.

For Indian Enterprises and Government Agencies

The sovereignty argument is real. Data location control, pricing in rupees, and a domestic provider with 5,100+ GPUs are meaningful advantages for organizations subject to Indian data protection regulations. E2E Networks’ IndiaAI Mission contract adds credibility. But you should ask about specific compliance certifications, SLA guarantees, and data residency documentation.

For AI Infrastructure Competitors

Together AI, Fireworks AI, and Replicate should note the sovereignty positioning. JarvisLabs is not competing on catalog breadth or inference speed — it is competing on geography, data control, and price. If Indian enterprises prioritize sovereignty, the US-based providers’ catalog advantage may not matter as much as expected. However, JarvisLabs’ five-model catalog is a significant limitation that competitors will exploit.

For E2E Networks Investors

The Managed Endpoints launch extends E2E Networks’ product portfolio beyond raw GPU rental into managed services. This is strategically important because managed services carry higher margins and create stickier customer relationships. Q1 FY27’s 75.2% EBITDA margin demonstrates the profitability potential of the GPU cloud business. The Managed Endpoints service, if adopted, could accelerate revenue growth beyond the 334% YoY rate seen in Q1.

But investors should watch the model catalog expansion. Five models is a starting point, not a competitive catalog. The speed of adding new models will determine whether Managed Endpoints becomes a meaningful revenue contributor or remains a niche offering.

What Is Genuinely New vs. What Is Repackaged

Genuinely New

The public disclosure of validation effort — 1,000+ GPU-hours and $5,000 per model, with benchmark replication scores — is genuinely new. Most managed inference providers treat their optimization work as proprietary. JarvisLabs’ transparency about the engineering behind each endpoint is a differentiator.

Repackaged

The concept of managed endpoints for open-weight models is not new. Together AI, Fireworks AI, Replicate, Groq, and Cerebras all offer variants of this service. JarvisLabs’ version adds sovereignty and localized pricing, but the core concept — pre-configured API endpoints for open-weight models — is well-established.

Unclear

Pricing model and rates. SLA guarantees. Catalog expansion timeline. Latency and throughput benchmarks. Data residency certifications. Customer adoption. Whether the five-model catalog will expand fast enough to compete with Together AI’s 200+ offerings.

The Bigger Picture

The JarvisLabs Managed Endpoints launch is a small but strategic move within a larger story. E2E Networks is building India’s sovereign AI infrastructure — 5,100+ GPUs, B200 clusters, IndiaAI Mission contracts, and an L&T partnership. The company’s revenue grew 334% year-on-year in Q1 FY27, with 75.2% EBITDA margins.

Managed Endpoints extends this infrastructure story into the application layer. Instead of renting GPUs and configuring inference software, enterprises get pre-validated endpoints. The value proposition is clear: cut weeks from deployment, pay in rupees, keep data in India.

But the competitive landscape is unforgiving. Together AI offers 200+ models with proven inference optimization.

Fireworks AI delivers the fastest inference on open-source models. Groq and Cerebras push hardware boundaries. Replicate covers 2,000+ models. JarvisLabs offers five.

The question is whether sovereignty, price, and validation transparency can overcome catalog breadth. For Indian enterprises subject to data protection regulations, the answer may be yes. For global enterprises with no sovereignty constraint, the answer is likely no — not with five models against 200+.

JarvisLabs plans to expand the catalog with speech-to-text, text-to-speech, and text-to-video workloads. The speed of that expansion will determine whether Managed Endpoints becomes a credible alternative to Together AI and Fireworks AI, or remains a niche offering for the Indian market.

Subramanian wrote on his blog: “I want to take Jarvis Labs into a major cloud provider in the world built from India.” Managed Endpoints is a step in that direction. But the distance between five models and a global inference platform is measured in engineering hours, GPU costs, and model partnerships — the same 15 to 20 working days per model that JarvisLabs is trying to eliminate for its customers.

GLM-5.2 validation: 1,000+ GPU-hours, $5,000 in compute, 80.9% on Terminal-Bench 2.1 vs 81.0% reported by Z.ai. This benchmark replication is publicly disclosed — a level of transparency not standard among managed inference providers.

Contact: techrecasteditor@gmail.com