Welcome to Ignition, Catalyst Investors’ briefing on what we’re seeing at the intersection of AI, robotics, software, and growth equity, combining the best of human and AI expertise. 

In this issue:

1. Nvidia’s strategic battle
2. Salesforce buys agentic CX platform Fin AI
3. The agentic customer service market map

01 — THE SIGNAL

The Fate of Nvidia

“All that time

I sat alone in my tower

You were just honing your powers

Now I can see it all (see it all)”

From “The Fate of Ophelia” by Taylor Swift

In March 2000, Cisco Systems was briefly the most valuable company on earth. It made the routers and switches that carried the internet during the dot com boom, and Wall Street decided that made it worth more than $500 billion. The stock had grown 3,800 percent in five years. Then the buildout slowed, a faster competitor showed up in Juniper, and Cisco fell 88 percent. It never traded at a market capitalization that high again.

Nvidia is the hardware darling of this decade’s infrastructure boom the way Cisco was the hardware darling of the last one. It makes the chips that train and run AI models. It is, again, the most valuable company on earth. And it is up more than a thousand percent since ChatGPT launched. The comparison writes itself, which is exactly why it is worth taking seriously instead of waving away.

The Moat Is CUDA, and It’s Why Training Doesn’t Move

Cisco’s moat was hardware performance, full stop. Once cheaper switches could do the same job, the moat was gone. Nvidia’s moat was never really the chip. It’s CUDA, the software layer released in 2006 that made Nvidia hardware the only place where millions of developers’ code runs. Nvidia reports no separate software revenue. It doesn’t need to. CUDA is baked into the 75 percent gross margin, twenty years of accumulated switching cost that Cisco never had. That’s why training workloads haven’t moved an inch, even as every major AI company mutters about wanting off Nvidia.

Inference Is the Bridge Over the Moat

Training needs the deep, CUDA-dependent software stack. Inference doesn’t, not nearly as much, and that gap is precisely where the escape attempts are concentrated. Google built the TPU (Broadcom-designed, TSMC-built) and just split it into a training chip and an inference chip. Amazon built Trainium in-house through Annapurna Labs and put it under Anthropic’s biggest training cluster. Microsoft built Maia, inference-only for now, co-designed with input from OpenAI, already running production GPT-5.2 traffic. OpenAI is building its own chip with Broadcom, the same partner sitting on Google’s side of the table too.

Nvidia’s answer wasn’t just to build faster. It bought Groq for roughly $20 billion, its largest deal ever, to bolt proven low-latency inference silicon onto its own rack-scale systems. That’s not the move of a company confident its organic inference response is fast enough on its own.

Then Nvidia Poured the Hot Oil. It Built a Model.

Nemotron is Nvidia’s own open-weight model family, and it doesn’t try to be the smartest model available. It doesn’t need to be. It’s built to be the fastest and cheapest, closer to the Chinese cost-performance curve of DeepSeek, Kimi, and Qwen than to the closed labs it competes with. Nemotron is trained in Nvidia’s own low-precision format and tuned for Nvidia’s own inference software, so even a free model, running on any hardware a customer chooses, works best on Nvidia’s stack. It’s the CUDA playbook, run one layer higher, aimed at the one part of the business where Nvidia’s grip was never absolute.

It already has a real customer, and a vocal one. Palantir CEO Alex Karp has been putting Nemotron in front of government and defense clients under the banner of “sovereign AI”: secure, air-gapped deployments that can’t touch a commercial API. He’s made the pitch on television, criticizing OpenAI and Anthropic in the same breath (“something has gone completely wrong”). Karp is talking his own book, obviously. But these are real deployments, in a segment where data sovereignty rules out the closed labs on principle, not just on price.

Commoditize Me, I Commoditize Thee

Every one of Nvidia’s biggest customers is racing to escape it. OpenAI is building its own chip. Anthropic is running on two rivals’ silicon. Google, Amazon, and Microsoft are all fielding their own accelerators, funded by the very revenue Nvidia’s GPUs generate for them. That’s the commoditize-the-supplier playbook, and it’s aimed squarely at Nvidia’s chips.

Nvidia’s response is to run the same play back at them, one layer up. If it can’t stop customers from diversifying away from its silicon, it can make sure the model sitting on top of whatever silicon they land on is a free one that happens to run best on Nvidia’s own stack. Commoditize the model layer, and the chip layer stays valuable no matter who wins underneath it. Arm the rebellion and own the arms dealer’s margin either way.

Cisco never had a second moat. Its grip was hardware, and once hardware got cheap, so did Cisco. CUDA gave Nvidia the first moat that Cisco never built. Nemotron is the bet that it can build a second one, faster than the industry can route around the first. Whether “runs best on Nvidia” turns out to be as sticky as “the only place your code runs” is the question, and nobody, least of all Nvidia, knows the answer yet. The battle is joined.

02 — STRATEGY CORNER

Salesforce Buys Fin: Reading Between the Agentic Lines

Liz Miller at Constellation Research looked at Salesforce’s agreement to acquire Fin (the customer-service AI agent formerly known as Intercom) for a reported $3.6 billion, weighing whether it signals a coherent agentic vision or gaps being patched. Read the full piece here.

Fin is not a rounding error. The AI agent alone has crossed $100 million in annualized revenue, growing roughly 350 percent year over year with around 8,000 businesses running it, inside a parent doing close to $400 million in total recurring revenue.

• The asset matters more than the logo. Fin’s real value is arguably less about its customer-experience agent business and more about its proprietary model, Fin Apex 1.0, which the company claims beats general-purpose models like GPT-5.4 and Sonnet 4.6 on hallucination rates and resolution rates for customer service specifically.

• Apex fits the open-weight-foundation playbook we’ve traced across the last two issues: a proprietary model built on an undisclosed open-weight base, the same move Palantir says it is making with its own deployments on Nemotron (discussed above). Application-layer companies increasingly fine-tune their own models rather than wrapping around a frontier model.

• This is the Issue #2 headless thesis playing out one layer up. The SaaS product is becoming a database with agents on top, whether self-developed, acquired like Fin, or built by the customer on Agentforce. Value migrates to whoever owns the data and the agent layer, not whoever wrote the original application.

• The optimist’s case is architectural. Miller sees Fin, layered on Contentful, as a path toward one customer agent replacing Salesforce’s historic cloud-by-function structure.

• The pessimist’s case: why keep buying agentic capability if Agentforce were already delivering? Salesforce’s execution risk is organizational, not technical, since the turf war between its sales, marketing, and service teams won’t dissolve on its own.

The Catalyst Take

First, Fin’s Apex extends a pattern from Issue #4 and the Nvidia piece above: the value in AI at the application layer is migrating from frontier-model wrappers to fine-tuned applications built on an open-weight foundation. Second, this confirms the headless thesis, where the SaaS product is becoming a database with agents on top. Third, expect more of this type of consolidation. For example, ServiceNow closed its Moveworks deal in December 2025 for the same reason Salesforce bought Fin: buying can be faster than building against the clock.

03 — MARKET MAP

Agentic Customer Service: Who’s Building What

The market splits into three tiers: platform incumbents bolting agentic capability onto an existing product, AI-native challengers built from scratch, and a fragmented layer of SMB-focused “AI receptionist” plays.

The Incumbents (Build, Then Buy)

• Salesforce (Agentforce + Fin): Running both plays at once, it has Agentforce for do-it-yourself agent building and Fin as a $3.6 billion acquired agentic layer with its own model and installed base. Unclear whether these merge into one product or run in parallel.

• Zendesk AI (+ Forethought): Acquired Forethought in March 2026, folding its capabilities into the core ticketing suite. Like Salesforce with Fin, Zendesk bought the agentic vendor rather than closing the gap organically. Zendesk has the largest app marketplace in the category, but the AI sits on top of a twenty-year-old ticketing architecture.

• Freshworks (Freddy AI): Similar incumbent position to Zendesk, adding agentic features to an existing ticket-and-seat business. More conservative on autonomous resolution claims than AI-native competitors, whether from caution or a weaker model.

• Kore.ai: The outlier, a horizontal enterprise conversational AI platform that competes in CX alongside employee-experience and IT-service use cases. Rarely the specialist choice for a buyer whose only problem is customer support.

The AI-Native Challengers

• Sierra AI: Valued around $15.8 billion after a $950 million round in May 2026, ARR near $150 million. High-touch, outcome-priced, positioned as a managed partner for the largest enterprises rather than a self-serve tool.

• Decagon: Valued at $4.5 billion after a $250 million round in January 2026, ARR around $35 million and growing nearly 300 percent year over year. Leans toward internet-native customers wanting more technical control than Sierra’s managed model offers. Sierra and Decagon are a speed-versus-depth tradeoff, not just a valuation gap.

• Ada: Same resolution-agent category as Decagon and Sierra, generally the more established, less richly funded alternative, differentiated more by integration breadth than architecture.

The AI Receptionist Layer

A more fragmented tier sits below the enterprise market: AI receptionists for SMBs like dentists, contractors, and dealerships. Companies like Retell AI and Bland AI sell voice infrastructure directly to SaaS companies, agencies, and technical operators, allowing them to build the voice layer directly into their products. Look for vertical SaaS companies like Weave Communications, ServiceTitan, and Toma to keep rolling out AI receptionist products that incorporate their deep domain expertise.

The Catalyst Take

Every incumbent is buying its way to agentic parity, and AI-native challengers prove a legacy platform isn’t required to reach billion-dollar valuations. None of this is surprising: CX, like coding, is one of the most obvious LLM use cases, which is why the category has consolidated fastest. Expect the SMB layer to consolidate around vertical solutions.

Sources: Constellation Research / Liz Miller (June 16, 2026), Sacra, Wikipedia, ServiceNow SEC 8-K, Crescendo.ai, Dynamic Business, eesel AI, Aircall, CloudTalk, Y Combinator company directory.

This newsletter is for informational purposes only and does not constitute investment advice or an offer to sell or a solicitation to buy any security.

Forward to a founder, operator, or investor who is navigating AI adoption.

ir@catalyst.com

261 5th Ave

Suite 1102

New York NY 10016