In this briefing
- 01Capital use begins to correspond to different layers of the enterprise AI delivery stack
- 02Cohere binds financing to local compute and geographic commitments
- 03FuriosaAI directs financing towards a production card and a next-generation chip
- 04Serval and Mintive direct funding towards business execution and runtime identity
- →What to watch next
- ↗Sources and verification
- Cohere completed an initial closing of its Series E financing, and Cohere stated a target of up to CAD 7 billion. The Government of Canada committed up to CAD 1.6 billion through a staged equity investment, but Cohere did not disclose the total amount currently committed or the timing of the final closing.
- FuriosaAI completed a $620 million Series D financing, and FuriosaAI says that the RNGD inference card is in mass production. FuriosaAI's expansion plan provides a supply signal, but vendor performance figures, conflicting power materials and unpublished pricing still require separate verification.
- Serval and Mintive are directing funding respectively towards deterministic enterprise workflows and Agent identity permissions. The customer, migration and deployment figures that the two companies published come mainly from company statements, and the financing amounts cannot prove long-term reliability or market advantage.
Capital use begins to correspond to different layers of the enterprise AI delivery stack
On 29 October, Cohere announced financing for sovereign models and compute infrastructure. Serval announced enterprise workflow automation financing on the same date, and Mintive also announced financing for Agent identity, permissions and tracing on the same date. On 28 October, FuriosaAI announced financing for inference-card mass production and next-generation chip development. The four transactions correspond respectively to model and deployment control, inference-hardware supply, business-process execution and runtime identity management.
The scale and status of the four transactions are not the same. Cohere announced a financing that has completed an initial closing but still states a maximum amount as a target. FuriosaAI and Serval announced completed financings, while Mintive announced cumulative financing and one new financing round within that total. The four companies also did not announce a shared product, interface or delivery arrangement, so the transactions show only that capital is covering different delivery layers; they do not prove that the four companies have formed a unified technology stack.
For enterprise adoption, the stated use of financing can indicate which type of capability a supplier is preparing to expand, but a financing amount cannot directly answer whether a product is available, whether systems are compatible or whether a service meets production requirements. An enterprise can convert a market signal into procurement evidence only when the enterprise maps model deployment, hardware supply, business execution and identity permissions separately to testable products, versions, regions, service levels and runtime records.
Cohere binds financing to local compute and geographic commitments
On 29 October, Cohere announced that it had completed an initial closing of its Series E financing, and Cohere stated a target of up to CAD 7 billion for the round. Cohere says that the target corresponds to a valuation of CAD 35.7 billion, and a Schwarz Group company is the lead investor. Reuters notes that Cohere did not disclose the total amount currently committed or the timing of the final closing, so the maximum CAD 7 billion remains a target rather than money that has all been received.
On the same date, the Government of Canada announced that it would make an equity investment of up to CAD 1.6 billion in Cohere on commercial terms. The Government of Canada says that the investment will proceed in stages, and the Government of Canada requires Cohere to add at least 1,000 Canadian technology jobs by 2030, invest at least CAD 1.7 billion in Canadian compute infrastructure and retain its Canadian incorporation and headquarters. The government announcement does not disclose the payment conditions for each stage, the ownership percentage or separate valuation terms.
For enterprise adoption, sovereign model supply involves not only model weights or data location, but also compute infrastructure, corporate control, deployment location and continuing operational capacity. Staged investment and geographic commitments can add traceable delivery conditions, but an enterprise still needs to check the specific model, private-deployment method, data processing, supported regions and service responsibilities; government ownership also cannot automatically prove that every customer environment meets regulatory requirements.
FuriosaAI directs financing towards a production card and a next-generation chip
On 28 October, FuriosaAI announced that it had completed a $620 million Series D financing at a post-transaction company valuation of $2.9 billion. FuriosaAI says that the RNGD inference card entered mass production in January 2026 (month 1) and that the company plans to produce 20,000 cards during 2026. FuriosaAI also says that the company has secured memory and foundry capacity so that production in 2027 can rise to more than twice the 2026 level. The third-generation chip remains in development, and the third-generation chip is planned to use a 2-nanometre multi-die design and HBM4 or 4E memory.
FuriosaAI's announcement on the same date describes RNGD as a 180W air-cooled PCIe card. Eight RNGD cards form an NXT RNGD Server with a 4U form factor, 3kW power, 384GB of HBM3 and 4 petaFLOPS of FP8 compute. An earlier data sheet that remains linked from the FuriosaAI website gives a 150W thermal design power for one card, and the earlier data sheet gives 48GB of HBM3 memory and 1.5TB/s of memory bandwidth for one card. The two official documents contain a power conflict, so the actual procurement configuration still needs confirmation against the latest model documentation.
FuriosaAI claims that RNGD can match the throughput of RTX PRO 6000 with the same card count and can serve twice as many users per kilowatt. FuriosaAI also says that SDK 2026.4 supports mixed batching, prefix caching and native Kubernetes, and FuriosaAI says that developers can deploy models through PyTorch. Those performance results come from vendor materials. The public announcement does not provide complete conditions for quantisation, context, batch size, concurrency or independent reproduction, and it does not provide product pricing, a specific delivery timetable, a network topology or a complete after-sales service level.
For enterprise adoption, mass-production status, capacity plans and the software stack are closer to purchasability than peak compute alone, but an enterprise still needs to place card-level power, server power, memory capacity, target models, framework versions, price, supply and support responsibilities in the same configuration review. The third-generation chip that remains in development can show the supplier's roadmap, but the third-generation chip cannot replace actual delivery of the current RNGD; vendor tests also cannot replace tests of time to first token, throughput, concurrency, energy use and failure recovery under the target workload.
Serval and Mintive direct funding towards business execution and runtime identity
On 29 October, Serval announced that it had completed a $150 million Series C financing at a transaction valuation of $3 billion. Serval says that Catalyst converts business requirements into deterministic tools and that administrators manage guardrails, permission checks and approvals. Reuters reports that Serval initially focused on information-technology work involving employee access, incident handling and software and hardware assets, and that customers still generally require human approval for password resets, application access and sensitive system changes.
On the same date, Mintive announced total financing of $11 million, including a latest financing round of $7 million. Mintive separates its product into Agent identity, a Model Context Protocol gateway, a large language model gateway, Agent monitoring and collaborative Agents. Mintive says that each Agent can receive scoped and revocable credentials, and that the monitoring system records the initiator, Agent, server, model, parameters and result. TechCrunch notes that the Agent management market already includes large cloud providers, model providers, monitoring providers and many start-ups.
Serval and Mintive both identify enterprise controls beyond model capability as product objects, but the two companies operate at different layers. Serval mainly converts business requirements into controlled workflows, while Mintive mainly provides identity, permissions and tracing foundations across different models and tools. The migration time, customer count and deployment effects that the two companies published come mainly from company statements, and the available materials do not provide unified independent success rates, failure rates, service levels or long-term renewal data.
For enterprise adoption, business execution requires repeatable tools and approval rules, while cross-model operation requires revocable identities and complete tracing. Capital entering these two layers can expand product supply, but an enterprise still needs to verify object identification, permission inheritance, human approval, fail-closed behaviour, log completeness and cross-system rollback. A larger number of market competitors can provide replacement choices, but a larger number of market competitors also raises the verification requirements for interface compatibility, supplier continuity and migration cost.
What to watch next
- Can Cohere disclose the amounts currently committed and received in the Series E financing, the final closing date and the specific verification conditions for the Government of Canada's staged investment?
- Can FuriosaAI publish consistent and current RNGD power documentation, formal pricing, regional availability and a complete network configuration, and can FuriosaAI provide independently reproducible model latency, throughput, concurrency and energy-use results?
- Can Serval and Mintive use comparable customer data to explain permission denials, human approvals, workflow failures, cross-system rollbacks and long-term migration outcomes?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing