Soket AI raises $4 million from Bajaj Finance and Pennant | Magica
Soket AI raises $4 million from Bajaj Finance and Pennant
Editorial Team
••📖5 min read
Bajaj Finance and its investee Pennant Technologies have invested roughly $2 million each in Soket AI. The backing connects Indian model research with potential financial-sector users, but costs, performance and production deployments remain undisclosed.
Bajaj Finance and its investee Pennant Technologies have invested roughly $2 million each in Soket AI.
The backers want affordable models they can operate within their own infrastructure; the announcement supplies no measured deployment results.
Soket has launched EKA-1, but larger models, broader language coverage and competitive performance remain research goals.
Soket AI, a Bengaluru-based research company developing Indian language models for mathematics, coding and reasoning, has secured roughly $4 million from Bajaj Finance and Pennant Technologies. The financing links its research to two connected businesses seeking AI they can use in financial services while retaining control over sensitive data.
Company-reported contributions to Soket AI: approximately $2 million from Bajaj Finance and $2 million from its investee Pennant Technologies. Source: Bajaj Finserv.
Capital from potential users
Bajaj Finance, a deposit-taking non-bank lender and subsidiary of financial-services group Bajaj Finserv, invested approximately $2 million through Finserv Intelligence, the group's applied-research initiative. Pennant Technologies, a lending-software provider to banks and financial institutions in which Bajaj Finance already holds an investment, contributed another $2 million, according to the November 18 announcement.
The investors bring potential users and financial-services expertise alongside money. Bajaj Finance says it serves more than 128.85 million customers across businesses including consumer loans, small-business finance and commercial lending. That is the lender's reported customer base, not adoption of Soket's technology.
Bajaj Finance says Soket's reasoning models could make advanced AI economically viable while allowing deployment on its own infrastructure, and that Indian-language capabilities could help it serve customers in their preferred languages. Pennant says it is building AI agents into its lending platform and wants locally developed models that regulated institutions can operate and govern themselves.
Those are the investors' expectations. The partnership is expected to create opportunities for applied research, enterprise applications and capability building; the announcement does not identify a completed production deployment or quantify savings.
Anthropic finished above its IPO price after a reported $81.35 billion offering, while US and Asian stocks fell. Its financing success leaves questions about computing costs and enterprise demand unresolved.
This is also part of a broader investment strategy. A report on the transaction says Bajaj Finance acquired a 5% stake in AI video-generation startup TrueFan AI two months earlier. Finserv Intelligence describes a five-to-ten-year horizon for impact, combining startup investments, academic collaborations and in-house specialists.
Company-reported: EKA-1 has 26 billion parameters; support for more than 100 languages is a Project EKA goal, not established EKA-1 coverage. Source: Bajaj Finserv.
A launched model and a larger unfinished program
The investment announcement says Soket has launched EKA-1, a 26-billion-parameter foundation model under Project EKA, its research program for sovereign AI—models developed in India that organizations can operate in controlled environments. The program focuses on mathematics, coding and reasoning.
Its roadmap extends to models with hundreds of billions to a trillion parameters and support for more than 100 languages, including Indian, Global South and programming languages. Those are program ambitions, rather than established capabilities of EKA-1. Parameter counts describe model scale; the announcement provides no comparative performance results.
Soket has also launched Loop, an open-source harness designed to execute long-running AI-agent tasks. The company is researching both models and systems for putting them to work.
Abhishek Upperwal, Soket's founder and chief executive, described the company's earlier Pragna-1B small language model as an experiment in building foundational models for Indic languages. In a March 8 interview, he said the company was building on that work with a planned 120-billion-parameter model supporting all 22 scheduled Indian languages. He also said it was open-sourcing parts of its datasets.
Upperwal then described a text-first strategy, with multimodal capabilities to follow. He expected the coding model to achieve performance comparable to Anthropic's Claude Code within about a year. That was his target against an existing coding tool, not a demonstrated like-for-like benchmark or a result established by the new financing.
He also reported interest from domestic IT services firms seeking data guardrails and said demand was strongest from government departments and public-sector entities. Discussions covered defence, agriculture and financial services. These were his accounts of potential demand and deals in progress, rather than confirmed deployments.
The November announcement identifies a launched 26-billion-parameter model but does not explain its relationship to the earlier 120-billion-parameter project or report completion of that larger model.
Government support brings resources and alternatives
Soket's government backing predates this round. The IndiaAI Mission is the government's program supporting domestic models, datasets and access to computing infrastructure. A July 30, 2025 ministry account listed Soket among four selected foundation-model developers and described its planned open-source 120-billion-parameter model for India's linguistic diversity.
Soket was one participant in that effort. The same account identified Sarvam AI's planned open-source 120-billion-parameter model, Gnani AI's 14-billion-parameter voice model and Gan AI's 70-billion-parameter multilingual model targeting text-to-speech. These projects addressed different uses and were described as under development, so their parameter counts do not establish a performance ranking or interchangeable substitutes. They do show that domestic model development is a broader field than this financing alone.
The ministry said the compute program supported 40% of GPU costs. This describes the scheme in July 2025, not a disclosed subsidy payment or training bill for Soket. It means the private round alone cannot establish the resources available to the research program.
Access also differs from usable capacity. In the March interview, Upperwal said IndiaAI infrastructure was improving after early GPU rollout delays. He estimated that around 6,000–8,000 H100 GPUs were usable at that time. That was his assessment of the wider infrastructure, not Soket's allocation or a current November capacity count. The financing announcement does not specify the compute required for EKA-1 or its larger successors.
The next decision turns on measured deployment results
The investors' interest makes the next test concrete: can Soket deliver useful financial-sector results at a sustainable cost, under the institutions' own data and governance controls?
Running models on an institution's infrastructure can give it control over where data is processed. It does not by itself establish accurate reasoning, safe operation or lower total costs. The announcement discloses no valuation, ownership percentages for this round, model-serving cost comparison or measured financial-services deployment results. It also supplies no dated milestones for the larger models.
The evidence needed next is performance on defined financial tasks, error rates and oversight requirements, infrastructure needs, and costs against alternatives serving the same workload. Those results would show whether the strategic backing can translate research into systems the investors can economically use.
Smith Point Capital led the financing, with customer GE Vernova also investing. BrightAI aims to connect equipment sensing with field repairs, but established maintenance tools set a tougher benchmark than its estimate of trillions of dollars in infrastructure waste.
Everful has raised $48.5 million to expand its AI-assisted factory sourcing and fulfillment business. Its pitch rests on integrating work that smaller retailers struggle to manage, though rivals offer overlapping services and its reported scale remains company-provided.
OpenAI is testing its persistent ChatGPT agents with selected Plus subscribers and expanding Business Standard access in the United States. Usage allowances remain unclear for the new tiers, while Enterprise customers face a January billing start.
The Senate voted 62–33 for BAAJA, pairing faster energy approvals with data-center transmission-cost rules. The House must still act, while opponents dispute environmental safeguards and which computing facilities would have to pay.
Anthropic’s debut put multibillion-dollar values on venture investors’ holdings, though estimates disagree. The offering also supplies about $79.4 billion in gross company proceeds as the Claude developer expands its computing infrastructure, while founders and a special trust retain unusual authority.
Only two submissions supported Western Downs Digital Park in two differing accounts of its consultation. The proposed A$32 billion campus has a conditional first-stage Anthropic lease, while water, power and local benefits remain contested.
Epic says it is using AI to rewrite parts of Unreal Engine and plans to let developers connect their preferred models to UE6. The strategy combines claimed maintenance gains with a broader overhaul of programming, distribution and portable game assets, while rival Unity already offers outside-model connections.
At Ignite, Nvidia and Microsoft are expanding local coding AI, cloud desktop streaming and Azure infrastructure. The announcements build on an existing Windows AI strategy, with AMD alternatives, staggered availability and no comparable deployment-cost results in the announcements.
Microsoft and Nvidia are expanding Azure’s Vera Rubin virtual-machine preview, with systems already running in its data centers. Microsoft’s separate confidential-AI showcase covers four GPUs, while the announcements leave pricing, regions and general availability unspecified.
Meta is reconsidering its 2030 climate deadline after net emissions reached 11.8 million tonnes in 2025. Hardware and infrastructure purchases dominate its reported footprint, while new gas generation adds a longer-term challenge.
Three House committee chairs want the Remote Access Security Act in the annual defense bill as scrutiny of Nvidia intensifies. The available Senate text would authorize controls on specified foreign users’ remote access, with computing limits left to Commerce.
Google Quantum AI reports tests of quantum contextuality spanning up to 105 qubits on Willow processors. The experiments expose hardware errors and beat restricted classical strategies, while the largest task relies on an unproven classical benchmark.
A bipartisan House disclosure shows requests to sell 3.4 million AI chips to China and approvals for more than 879,000. The oversight question is whether Commerce verified the buyers, supply protections and access restrictions behind those authorizations.
OpenAI says its internal Performance Factory tracked more than 1,700 investigations in just over three months. About 400 cases reached proposed fixes, but far fewer shipped repairs used Codex’s code unchanged, and engineers retain approval.
Activists protested Anthropic’s stock-market debut outside Nasdaq, with at least one arrest and conflicting accounts of the total. Their campaign targets retirement-fund exposure to AI shares, alongside concerns about jobs, data centers and public accountability.
Anthropic says cash, growing revenue and multiyear payment schedules support its vast compute commitments. Its filing shows why flexibility matters: short-term capacity costs more, while major cloud contracts require payment regardless of usage.
OpenAI is rolling out live transcription for dictated ChatGPT prompts across free and paid plans. Users can see words as they speak and review the text before sending, but the company has not supplied performance measurements or a replacement recording maximum.
Boltz has raised a Bessemer-led $110 million Series A to expand molecular-design tools sold by usage. Its open-source research brought broad adoption; the commercial test is whether paid computation delivers repeatable gains in laboratory discovery.
Donald Trump said he and Xi Jinping supported minimal AI guardrails but had not discussed definitions. Their Shenzhen meeting leaves unresolved how bilateral safety cooperation would coexist with U.S. opposition to global regulation.