Two announcements landed five days apart. Vesta raised $30 million on October 8 to grow its AI-powered loan origination system. TD expanded agentic AI across its mortgage operations on October 13, the Tuesday after the Columbus Day bank holiday. Underneath both, a dependency that has been building all year came into sharper focus. Valon, which raised $150 million earlier in 2026, now runs all of Newrez's loan servicing. The concentration is deep enough that Newrez's parent company flagged it as a risk in an SEC filing.
Read together, the three companies give three different answers to one question: who holds the operational memory of the mortgage business, and what happens when the relationship changes?
Mortgage work happens in two stages. Origination takes the application, verifies income, underwrites the risk and closes the loan. Servicing is everything after that: collecting payments, managing escrow (the funds held for taxes and insurance), handling delinquencies, answering borrower calls. The two stages run on separate software, sold to separate buyers, with separate economics. Vesta targets origination. Valon targets servicing. TD builds both itself. Each choice produces a different dependency and a different answer to who keeps institutional capability over time.
| Vesta | Valon | TD | |
|---|---|---|---|
| Workflow position | Origination system of record | Servicing system of record | Both, built internally |
| Buyer relationship | Replaces lender's LOS | Replaces servicer's platform | No external buyer |
| Who holds data | Lender owns loans; Vesta processes | Operational history accumulates on Valon's platform | Bank owns everything |
| Switching cost trajectory | High upfront (system migration), stable over time | Compounds annually as institutional memory deepens | N/A — cost is ongoing headcount |
| Evidence quality | Customer-investor commitments; no independent performance data | SEC-filed risk disclosure; production migration underway | Company-reported metrics; no disclosed methodology |
Vesta: new origination system, customers as investors
Vesta replaces the lender's loan origination system, the core software where applications arrive, documents get processed and underwriting decisions get made. Its agents read documents, extract data, check compliance and follow up with borrowers. Vesta says agents and automated workflows handle about 40% of tasks on its platform. It hasn't disclosed how it measures that, and I found no independent production data that corroborates it. Treat the number as a company claim.
On architecture, Vesta has announced a partnership with Blend that splits origination into two layers. Blend runs the borrower-facing point-of-sale and pre-underwriting; Vesta runs fulfillment and holds the system of record. Read the partnership as an ecosystem signal. The likely interpretation is that Blend looked at the full LOS-replacement market and concluded Vesta's position was better worked through integration than fought head-on. TechCrunch reported that Vesta logs agent actions and reasoning for audit, and that lenders can begin with human approval and widen agent autonomy later. What isn't public is the architecture underneath: how deep the trace logging goes, who governs compliance determinations, how audit records are structured. That limits how far anyone outside can evaluate the product's trust design.
The round carries its own signal. Three of the investors, Pennymac, New American Funding and nbkc bank, are also Vesta customers. Pennymac adopted Vesta in 2025 and was already live. New American Funding plans a phased migration in 2027. nbkc confirmed both its investment and its adoption on October 13. The other investors (Citi Ventures, FirstKey Mortgage, First American's Parker89) are mortgage-industry players, but I found no public evidence that any of them is a Vesta customer.
Customer-investors aren't unusual in mortgage tech. D.R. Horton invested in Tidalwave while becoming a customer. Mr. Cooper traded servicing IP to Sagent for a 19% equity stake and became Sagent's largest customer. The structure ties financing to distribution: the startup gets committed volume, and the customer gets equity upside plus some say over the roadmap. Whether Vesta's investors agreed to exclusivity, minimum loan volumes or referral obligations is not disclosed.
Dependency profile for the buyer: High switching cost, because you're replacing your origination system and that is a multi-year migration. But the lender still owns the loans and the borrower relationships. Vesta processes; the lender decides. What's at risk is operational efficiency. The underlying business data stays with the lender.
Valon: the platform that becomes the servicer's memory
Valon replaces the servicing system, the software that tracks every payment, escrow calculation, borrower interaction and delinquency action once the loan closes.
The strongest evidence here is an SEC filing. Rithm Capital's 2025 annual report discloses that its subsidiary Newrez is migrating all loan servicing onto Valon's platform, and it lists vendor concentration, meaning reliance on a single technology supplier, as an operational risk. That is a publicly traded company telling its shareholders, in an audited document, that it sees the dependency and is going ahead.
An SEC-disclosed vendor-concentration risk, filed by a company proceeding with full migration anyway, is the clearest evidence in mortgage AI that a vertical platform can become the place where an institution keeps its operational memory.
Why go ahead? Because servicing is where operational memory accumulates. Payment histories, escrow adjustments, loss mitigation records and borrower correspondence add up to years of institutional knowledge, and that knowledge sits wherever the system of record sits. Some of it is legal obligation. A servicer that inherits a loan inherits the promises made to the borrower, including forbearance agreements, loss mitigation terms and modification discussions still in progress. The platform has to encode those commitments along with the payment history. Once Valon's platform is the system of record, Valon holds the servicer's institutional memory, including obligations the servicer is legally required to honor.
Federal regulators recognize the problem in principle. Interagency guidance from 2023 says financial institutions should keep exit plans, data portability rights and the resources to move work to another provider. Regulation X requires servicers to hand over loan information in usable form during servicing transfers and to keep accurate records of what service providers do. But the guidance says explicitly that it creates no new legal requirements. And FFIEC examination guidance names the gap directly: after years of outsourcing a function, a company may no longer have the staff or internal capability to bring it back. The rules require a plan, with no assurance the institution can carry it out.
Dependency profile for the buyer: The deepest of the three. The servicer's operational knowledge moves into Valon's system over time. Switching costs compound with each year on the platform, driven by the servicer's eroding ability to operate on its own rather than by contractual lock-in. Rithm's filing suggests it understands the tradeoff and has judged the platform's value to outweigh the concentration risk.
TD: build internally, keep the capability
Of the three, TD's position looks least like a startup story and most like an institutional bet.
The bank acquired Layer 6, a 17-person AI research company in Toronto, in 2018. Layer 6 now has more than 200 people. TD's 2025 annual report counts over 2,500 AI-related roles across the bank (data scientists, ML engineers, specialists) and names TD AI Prism, a proprietary predictive foundation model, as part of its platform work. The mortgage expansion announced October 13 is one piece of a broader deployment covering investing, insurance, call centers, branches, wealth management, small-business banking and credit-card approvals. As of August 2025, TD had launched more than 60 AI applications.
TD doesn't build all of it. In September it announced a $25 million, three-year partnership with Cohere, whose engineers will work alongside Layer 6 on enterprise AI applications. TD owns the team, the data, the deployment authority and the policy layer, and uses outside model infrastructure. How the mortgage agents are built isn't public, including whether they share infrastructure with the credit-card underwriting models or the contact-center assistants and what the trust and auditability layer looks like. TD reports faster processing times and higher approval rates. Those figures are self-reported, with no disclosed methodology.
Dependency profile for the buyer: There is no external buyer. TD builds the AI around its own policies, data and compliance requirements, and it keeps the institutional capability because it never outsourced it. The cost is headcount. 2,500 AI-related roles is a commitment most lenders can't match, which leaves an open question: is this path a structural advantage only the largest banks can use, or a model a smaller lender could follow at its own scale?
Google: horizontal platform, no mortgage product
In a prior piece on legal AI, I described Google's approach as owning the platform while renting the vertical. In mortgage, Google has nothing to rent yet. Gemini Enterprise for Financial Services, in early access since August, is aimed at capital markets and corporate banking and connects to data sources like FactSet, LSEG and S&P Global. I found no connections to loan origination or servicing systems, no Fannie Mae or Freddie Mac integrations, and no disclosed mortgage customers. A lender could build on Google's infrastructure, but mortgage AI needs three things a horizontal platform doesn't supply: system-of-record status, where loan data lives authoritatively; GSE-compliant underwriting workflows; and the regulatory integration that makes a servicing platform legally defensible. Keep watching for changes. For now, Google isn't in this market.
The incumbent that sets the ceiling
ICE Mortgage Technology is the company most obviously missing from the startup framing. Its Encompass platform is the most widely used loan origination system in the U.S., and its MSP platform is a major servicing system. ICE's 2025 annual report says it is adding AI agents to both.
That puts ICE up against Vesta and Valon at once, with something neither startup can offer: one vendor for the loan's whole lifecycle. If the incumbent adds good-enough AI to software lenders already run, the case for ripping out Encompass in favor of Vesta gets harder to make. Other companies hold narrower positions. Sagent competes with Valon in servicing. Tidalwave and Polly cover origination and capital-markets segments. Maxwell, which PLACE acquired in September and which claims more than 400 financial-institution customers, pairs origination software with operational services.
What to watch next
Valon/Newrez carries the most structural weight. It is the one development here backed by an audited regulatory filing. Watch whether Rithm discloses specific exit-planning measures in later filings, and whether other servicers follow Newrez onto Valon or read the filing as a warning.
Vesta needs independent production data next. The customer-investor structure says a lot about distribution and nothing about performance at scale. If Pennymac or New American Funding publishes loan-processing metrics attributable to Vesta, the evidence moves from company-claimed to independently corroborated.
ICE sets the ceiling. How quickly ICE ships AI into Encompass and MSP, and whether lenders consider it sufficient, determines the upper bound on what Vesta and Valon can capture.
Directional view, moderate confidence: This vertical is splitting in two. Large banks with the resources to build, TD being the clearest example, will keep AI capability in-house, where it serves as a competitive advantage and removes vendor-concentration risk at the same time. Lenders and servicers below that resource threshold face growing platform dependency, because the switching costs of vertical AI systems compound over time in a way traditional software licensing did not. I'd treat Valon/Newrez as the likely equilibrium for mid-size servicers who can't staff their own AI organizations, rather than an outlier. ICE is the variable that could break the split. If it ships AI that's good enough across origination and servicing, mid-size lenders get a third option: stay on the system they already use and let the incumbent close the gap. That option is open now. Whether ICE moves fast enough to keep it open is the most consequential question in mortgage AI over the next twelve months.
- CFPB on servicing transfers: The Bureau's supervisory findings on mortgage servicing transfers document cases where servicers lost pending modifications, mishandled escrow data, and forced borrowers to restart loss-mitigation applications — the concrete borrower-level cost of the migration risk this piece describes at the institutional level.
- Schneider-PTC and lifecycle control: Schneider Electric's $23.7 billion agreement to acquire PTC joins industrial automation software with CAD, PLM, and service-lifecycle products, making it the largest current test of whether connecting design records to operational systems creates a durable AI control point in field and industrial verticals.
- Sagent's Mr. Cooper precedent: The 2022 transaction in which Mr. Cooper transferred servicing IP to Sagent for a 19% equity stake while becoming its largest customer is the closest structural precedent for Valon/Newrez — worth tracking whether that dependency has deepened or whether Mr. Cooper has maintained more operational independence than Newrez appears to be retaining.
- Entry-level displacement signal: A Stanford Digital Economy Lab working paper using ADP payroll data finds employment among workers aged 22–25 in highly AI-exposed occupations running 19% below less-exposed peers, with no comparable gap for experienced workers — a pattern that could reshape the pipeline of mortgage professionals available to staff the institutional capability this piece argues is at risk.
