DeepSeek V4 Opens the Frontier, Robinhood Bets on OpenAI, and BofA Gives 18,000 Advisors Their Hours Back

Three stories broke across the finance-AI stack this week that, taken together, tell you something important about where this moment is heading. The most powerful open-weight model in history just dropped at fourteen cents per million tokens. A retail brokerage bet $75 million on the company building the closed-weight alternative while simultaneously giving its own customers an AI agent that trades in plain English. And one of the largest wealth management firms in the world just handed its entire advisor workforce up to four hours back per client meeting.

These aren’t isolated product announcements. They’re three simultaneous bets on the same underlying thesis: frontier AI is now cheap enough, capable enough, and trustworthy enough to deploy at the point of actual financial decisions — whether you’re a hedge fund, a retail investor, or a Merrill Lynch advisor prepping for a portfolio review.

DeepSeek V4: The 1.6-Trillion-Parameter Open Question

On April 24, exactly one year after DeepSeek’s R1 release upended the assumption that frontier AI required US-controlled compute, the Chinese AI lab dropped V4 — and it’s a bigger swing than R1 was.

DeepSeek V4 comes in two configurations. The Flash variant weighs in at 284 billion parameters; the Pro variant runs 1.6 trillion total parameters with 49 billion active per forward pass through a mixture-of-experts architecture. That makes V4 Pro the largest open-weight model available by a significant margin. For context, the active parameter count is still lean enough to run efficiently — the MoE architecture means you’re not firing 1.6 trillion neurons per token, just the 49 billion relevant experts for each query.

The headline technical claim is the context window. V4 was built from the ground up around one million tokens as the default, not as a bolt-on capability. DeepSeek describes its Hybrid Attention Architecture — combining different attention mechanisms optimized for different positional distances — as what makes that default context window coherent rather than just technically possible. Most models that advertise million-token contexts quietly degrade in quality beyond 200–300K tokens. Whether V4 actually holds coherence at 1M is something the independent evaluations will determine, but the architectural commitment is notable.

The benchmark claim is bold: V4-Pro-Max reportedly outperforms GPT-5.2 and Gemini 3.0 Pro on some reasoning tasks, and the Codeforces rating of 3,206 places it 23rd among human competitive programmers. That’s a number worth watching as the hardened evaluation community picks it apart.

The compute story is what makes this geopolitically significant. DeepSeek built V4 on Huawei’s Ascend 950 chips, using Huawei’s “Supernode” cluster technology. One year after US export controls were supposed to cap China’s AI compute ceiling, China’s leading AI lab has released the world’s largest open model on domestic hardware. The export control thesis — restrict chips, slow the race — is empirically in trouble.

The pricing is the practical punchline. V4 Flash costs $0.14 per million input tokens and $0.28 per million output. V4 Pro costs $0.145 input and $3.48 output. At those rates, a financial services firm running tens of millions of document analysis queries per month is looking at sub-$20,000 monthly compute costs for frontier-class reasoning. That changes the ROI math on AI deployment for mid-tier banks, insurance companies, and asset managers who couldn’t previously justify frontier model API costs at scale.

Robinhood Goes All-In on Both Sides of the Bet

The same week DeepSeek dropped its open-weight bomb, Robinhood made two moves that look contradictory until you understand the strategy.

Move one: Robinhood Ventures Fund I announced a $75 million investment in OpenAI — the maker of the closed-weight GPT series that DeepSeek is specifically trying to displace. Robinhood is now a financial stakeholder in the leading proprietary AI lab, just as the open-weight challengers are becoming credible.

Move two: Robinhood launched Cortex, its own AI trading agent for retail customers. Cortex is built on a foundation of real-time market data, Wall Street-grade analyst reports, and research inputs that retail investors have never had direct access to before. The key capability is agentic: customers can tell Cortex in plain English to buy or sell equities and crypto, conduct market research on a specific position, and adjust account settings — while keeping the human in the confirmation loop.

The personalization layer is what distinguishes Cortex from a generic chatbot with a brokerage API. Cortex connects real-time news, market movements, and analyst signals specifically to what’s in your portfolio. When your holdings in semiconductor ETFs move after a DeepSeek announcement, Cortex surfaces the connection and explains the implication before you have to ask.

Robinhood Gold subscribers get early access. The full rollout is phased.

Read together, the OpenAI investment and Cortex launch reveal the underlying logic. Robinhood is betting that the AI powering retail investment decisions needs to be best-in-class — and in 2026, that still means OpenAI’s frontier models under the hood. The $75 million stake is partly commercial access, partly hedge, and partly a signal to institutional partners that Robinhood is serious about the AI-powered finance infrastructure layer. Cortex is the product that justifies all of it.

The long-term competitive implication is significant. If retail investors get access to research-quality AI analysis personalized to their own portfolios, the information asymmetry that defined the divide between retail and institutional investors for decades narrows considerably. That’s good for markets in the long run. In the short run, it’s a stress test for brokerages that haven’t invested in their AI product layer.

Bank of America Gives Back Four Hours Per Meeting

While DeepSeek and Robinhood grabbed the headlines, Bank of America’s Merrill Wealth Management quietly rolled out something that may have larger aggregate economic impact than either: Meeting Journey, now deployed to all 18,000 financial advisors across Merrill and Bank of America Private Bank.

Meeting Journey operates across three phases of the client relationship cycle. In preparation, it searches and consolidates client relationship history, recent account activity, and relevant market context into meeting-ready materials in minutes — the work that previously required advisors to manually pull data from multiple systems the morning before each appointment. During meetings, it functions as an AI notetaker (with client consent on virtual calls), capturing discussion highlights and generating a shareable summary. Post-meeting, it synthesizes decisions and next steps into tasks and follow-up documentation.

The time savings figure — up to four hours per meeting — is the one that should make every financial services leader stop and do the math. Merrill’s 18,000 advisors each conduct dozens of significant client meetings per month. At four hours per meeting across millions of annual interactions, the aggregate productivity return is in the hundreds of millions of hours annually across the advisory workforce. That time is now available for what advisors are actually paid to do: strategic planning, relationship depth, and the kind of judgment that an AI meeting tool explicitly cannot provide.

The BofA announcement coincided with a record Q1 2026 result: $8.6 billion net income, earnings per share of $1.11 — the highest EPS level in nearly two decades. Investment banking fees climbed 21% to $1.8 billion. The bank’s CFO explicitly cited AI productivity as a contributing factor to the operational leverage behind the numbers.

This is the enterprise AI deployment pattern that the McKinsey and PwC data has been pointing at for two years: not AI replacing advisors, but AI handling the preparation, documentation, and administrative overhead that consumes advisor time so that advisors can spend more of that time on the work that actually drives revenue and retention.

The OpenAI Paradox: $25B Revenue, $14B Loss, $852B Valuation

No finance-AI week is complete without checking in on the company at the center of all of it. OpenAI closed Q1 2026 with $25 billion in annualized revenue — up from $21.4 billion at year-end 2025 and roughly $6 billion eighteen months ago. The growth rate is genuinely extraordinary: one of the fastest revenue ramps in corporate history.

The funding round closed on March 31st: $122 billion at an $852 billion post-money valuation. Amazon committed $50 billion (with $35 billion contingent on an IPO or AGI milestone by December 2028). NVIDIA invested $30 billion. SoftBank contributed $30 billion. The list reads like a who’s who of the firms most dependent on OpenAI’s success — and most incentivized to ensure that success.

The number that doesn’t get enough attention: OpenAI is projecting approximately $14 billion in losses on $25 billion of revenue this year. Every dollar of revenue costs more than a dollar to generate, primarily in compute. The company is burning cash at a rate that requires either continued mega-rounds or an IPO, and the IPO prep is now visible: a new Head of Investor Relations hired, internal H2 2026 filing targets discussed, and a potential $1 trillion valuation floated in analyst models.

The paradox the numbers create: OpenAI is simultaneously the most valuable private company in history, the fastest-growing software business ever, and deeply unprofitable at scale. The bet embedded in that $852 billion valuation is that inference costs fall fast enough, and demand grows large enough, that the unit economics flip before the cash runs out. With Amazon’s contingent commitment tied to an IPO timeline, there’s structural pressure to list before December 2028 regardless of where profitability stands.

For financial services firms — who are simultaneously OpenAI’s fastest-growing customer segment and, in some cases, its investors — this creates an interesting dependency to manage. Building enterprise workflows on a company that loses $14 billion per year is a vendor risk calculation worth having explicitly.

The Pattern Underneath the Stories

Step back from the individual announcements and a coherent picture emerges. The finance-AI stack is being built on three simultaneous foundations that are mutually reinforcing.

Open-weight commoditization from DeepSeek V4 is driving inference costs to near-zero for frontier-class reasoning. This removes the compute cost barrier for mid-tier financial services firms that previously couldn’t justify frontier model deployment at scale. The competitive moat of the large proprietary model providers is narrowing.

Consumer-facing agentic finance from Robinhood Cortex and similar tools is compressing the information asymmetry between retail and institutional investors. When retail investors can access analyst-grade research personalized to their portfolios, the structural edge of institutional information access weakens.

Advisor productivity tooling from BofA Meeting Journey and its equivalents at JPMorgan, Goldman, and others is redirecting the labor economics of wealth management. Advisors who embrace AI preparation and documentation tools have more time for relationship depth. The advisors who resist it face a structural productivity disadvantage against peers who’ve adopted it.

None of these trends are reversing. The question for every financial services firm in 2026 isn’t whether AI will reshape the economics of their business. It’s whether they’re the ones doing the reshaping or the ones being reshaped by competitors who moved faster.

What to Watch

DeepSeek V4 independent benchmarks: The company’s claims about V4-Pro-Max outperforming GPT-5.2 and Gemini 3.0 Pro need independent verification. Watch for evaluations from the academic community and third-party benchmarking services in the coming weeks.

Robinhood Cortex Gold rollout: The initial rollout is limited to Gold subscribers. Watch user adoption rates and, more importantly, trading behavior changes among Cortex users versus non-users. If Cortex-assisted portfolios outperform non-assisted ones on risk-adjusted returns, the entire retail brokerage industry will be in a product sprint.

OpenAI IPO timeline: The $122B round with Amazon’s contingent commitment creates a structural forcing function toward a 2026 H2 filing. A public OpenAI would require financial transparency that doesn’t currently exist — and the disclosed loss figures would be a significant market event.

BofA Meeting Journey expansion: The tool is currently deployed to wealth management. Expansion to commercial banking, investment banking, and corporate banking would multiply the productivity impact by an order of magnitude. Watch the bank’s Q2 and Q3 operational metrics for signals.

US-China AI compute independence: DeepSeek’s V4 achievement on Huawei chips is a data point that will influence the next round of export control policy debate. If domestic Chinese compute can produce models competitive with US-hardware-trained frontier models, the effectiveness of chip controls as an AI governance mechanism is materially in question.

---

Sources

• DeepSeek Unveils Newest Flagship AI Model — Bloomberg
• Three reasons why DeepSeek’s new model matters — MIT Technology Review
• DeepSeek previews new AI model that ‘closes the gap’ with frontier models — TechCrunch
• DeepSeek V4 Released: Everything You Need to Know — FelloAI
• Robinhood Ventures Fund I Announces $75M Investment in OpenAI — Globe Newswire
• Introducing Robinhood Cortex, Strategies, and Banking — Robinhood Newsroom
• Merrill and Bank of America Private Bank Launch AI-Powered Meeting Journey — BofA Newsroom
• Bank of America’s 18,000 financial advisors just got a new AI tool — Fortune
• OpenAI tops $25 billion in annualized revenue — Domain-b
• OpenAI raises $122 billion to accelerate the next phase of AI — OpenAI
• Here Are 3 Financial Stocks Making Big AI Moves in 2026 — Motley Fool