AI Now: Elon’s $1T Package, Apple’s $600B for Trump & How Small Startups Win w/ Dave, AWG & Blitzy
Summary
The panel treats Elon Musk’s potential $1 trillion Tesla award as rational only because it is conditional on creating vastly more value. Musk would have to help take Tesla to an $8 trillion valuation, while his role combines CEO, chief marketer, and attention engine; Diamandis notes that conventional automakers spend roughly 7% of revenue on marketing while Tesla spends “zero.” The broader founder lesson is that communication has become part of the product, although “it’s going to change again and it’s going to change again.”
AI’s trillion-dollar capital cycle now needs labor automation and transformative science to generate returns commensurate with the spending. The panel recounts a dinner where a $600 billion U.S. investment commitment was made, with Dave Blundin saying the clips are unclear on whether Tim Cook or Mark Zuckerberg went first and that Zuckerberg matched it, while citing roughly $119 billion of additional OpenAI investment through 2029 and earlier projections of $5 trillion-$7 trillion for AI chips and related infrastructure. Alexander Wissner-Gross’s logic: first drive classes of labor and services toward negligible cost, then produce discoveries capable of justifying the capex.
Abundance would devalue money without eliminating scarcity; it would merely move the bottleneck. Peter Diamandis imagines molecular assemblers making an electric Ferrari at near-zero marginal cost, while Wissner-Gross counters that Star Trek has replicators but limited access to interstellar travel. His defining question is “what remains scarce” when energy and intelligence approach zero cost—a useful warning against assuming every constrained asset disappears together.
Mercor is the panel’s evidence that extreme AI valuations can still follow operating performance rather than pure speculation. Blundin says the company moved from an initial valuation near $30 million to $10 billion in two years while reaching a $500 million revenue run rate; its founders began at 18. His venture rule is to seek “undervalued, underappreciated talent,” with the wider shift being that founders aged 20-23 can now attack markets previously reserved for far more experienced teams.
Blitzy’s enterprise bet is that generating code is becoming a commodity, but understanding and safely transforming enormous codebases is not. Its platform claims support for more than 100 million lines of code, has successfully onboarded a roughly 60-million-line repository, and runs jobs from 12 hours to multiple weeks. The output is intended to arrive prevalidated, precompiled, and pretested because “the other side of a pull request” is expensive human labor.
Blitzy reports 86.8% on SWE-bench Verified versus the filmed leaderboard’s 75.2%, but the larger signal is that the benchmark may now be exhausted. The company says the result is reproducible through the SWE-bench CLI without benchmark-specific scaffolding, after processing 500 branches representing roughly 400 million ingested lines. Wissner-Gross argues that 86.8% is effectively near 100% because much of the remainder is flawed, creating demand for enterprise-scale tests against repositories such as Linux and VS Code.
The startup playbook is to become a large customer of the frontier labs, go deep into a problem they cannot operationalize, and improve whenever their models improve. Blitzy orchestrates Gemini, Anthropic, and OpenAI models rather than training a frontier model; Diamandis frames the labs’ spending as “a trillion dollars of R&D for Blitzy.” Its present enterprise claim is more grounded than raw code-generation hype: about 80% of the work automated on suitable projects and roughly 5x end-to-end development velocity, with humans receiving the remaining tasks explicitly identified.
Deep dive
1. A $1 trillion award prices Musk as both operator and distribution
Diamandis opens with the headline, while Blundin explains the conditionality: Musk could receive roughly $1 trillion in Tesla stock only if the company reaches benchmarks including an $8 trillion valuation—which Blundin describes as doubling the size of Microsoft and Nvidia. His defense is simple: “It’s just a fraction of what you create.”
Diamandis argues Musk has changed the CEO’s economic function. He is “not just the leader of the company, he’s the marketing voice”; where typical automakers might spend roughly 7% of revenue on marketing, Tesla spends zero because Musk personally generates attention, momentum, and customer demand.
Diamandis applies that observation to venture selection: a founding CEO who is “very, very shy” may struggle because investors now need communicators who can transmit conviction publicly. He adds the caveat that copying Musk is no permanent formula—media is changing, and the CEO archetype “is going to change again and it’s going to change again.”
2. Trillion-dollar capex requires trillion-dollar economic consequences
At the Trump dinner discussed by the panel, a $600 billion U.S. investment commitment was reportedly made; Blundin thinks Tim Cook went first but says the clips are unclear because they were cut and mingled, after which Zuckerberg matched it. Blundin’s comic reconstruction: it resembled a fundraiser while a Silicon Valley CFO wondered, “What the hell did he just commit to?”
The panel also cites something near $119 billion of additional OpenAI investment through 2029. Wissner-Gross recalls that projections of $5 trillion-$7 trillion for AI chips and related spending on fabs, data centers, and energy sounded absurd roughly 18 months earlier, yet spending on that scale had already become “entirely plausible.”
Wissner-Gross follows the capital to its required payoff: markets investing trillions will expect enormous revenue, plausibly from automating whole categories of labor or services and driving their costs toward zero. After that, AI may need to deliver transformative inventions and scientific discoveries; otherwise, he asks, “Why invest trillions in this?”
That leads to the abundance dispute. Diamandis imagines molecular assemblers turning energy, soil, titanium, and lithium into an electric Ferrari at near-zero marginal cost; Wissner-Gross answers with Star Trek, where everyone has replicators but not everyone travels between stars. Abundance does not end scarcity—it changes “what remains scarce.”
3. Mercor makes the case for backing exceptional talent earlier
Blundin presents Mercor as a valuation story supported by operating acceleration: approximately $500 million of revenue run rate within two years, a $10 billion valuation offer, and an initial financing valuation near $30 million. “I don’t want people to feel like that’s a bubble,” he says, because the revenue growth also broke precedent.
His selection principle was “undervalued, underappreciated talent and not so much concepts,” although Mercor already had the right concept. The unusual part was backing founders who were 18, had completed roughly one year of college, and became frustrated with its pace.
Diamandis contrasts today’s 20-to-23-year-old founders with the mid-30s average he recalls for venture-backed unicorn founders a decade earlier. Wissner-Gross expects educational incentives to distort further: if students believe advanced AI is imminent, accumulating credits may compare poorly with building a company immediately.
Diamandis invokes research suggesting Nobel-winning work often occurs in the first half of a researcher’s twenties. Wissner-Gross resists making that forecastable: pure AI or human-AI hybrids may soon perform much of the innovation, making age-productivity statistics “how things used to be at best.”
4. A bakery app exposed the bottleneck that became Blitzy
Elliott and Pardeshi met at Harvard Business School after very different operating careers: Elliott came through West Point and military service, while Pardeshi spent time at NVIDIA and had worked on generative models since Attention Is All You Need. Their friendship became unusually close; Pardeshi is godfather to Elliott’s youngest son.
During the GPT-3-to-3.5 era, a pro bono project for a Boston bakery supplied the spark. The bakery expected to spend $300,000-$400,000 on a mobile app; Elliott and Pardeshi built it overnight or over the weekend using a workflow now called vibe coding. “We acted like it took longer,” Elliott jokes.
Their key observation was that the humans had become the bottleneck: they repeatedly fed errors back into one model, handed the output to another, and iterated toward compiling, tested code. The founding hypothesis was to automate that multi-model refinement and remove commoditized development labor from the loop.
Today Elliott defines Blitzy as an “enterprise-grade autonomous software development platform.” It absorbs a codebase, converts requested work—such as COBOL-to-Java migration or steady-state development—into long-running inference jobs, and aims to return high-quality code that has already been validated, compiled, and tested.
5. Blitzy goes after codebases too large and old for humans to comprehend
Financial institutions may still depend on COBOL, PL/I, and other decades-old systems that work but are frightening to modify. The people who understand them are disappearing, wholesale replacement has historically been uneconomic, and management often lacks visibility into what tens of millions of lines actually do.
Blitzy first indexes the source and gives the enterprise a functional map, then uses that context for migrations or new features. Blundin calls the experience of asking 10 million undocumented lines to explain themselves in plain English—and identify bugs—“mind-blowing.”
Elliott says 20-million-line repositories are common, with roughly 60 million lines the largest successfully onboarded; the platform’s claimed context capacity exceeds 100 million. Jobs have run for 12 hours and, on massive codebases, multiple weeks—an engineering regime intentionally separated from interactive copilots.
That distinction is Blundin’s investor thesis. Cursor, Windsurf, Replit, and Lovable excel at rapid interaction, but an agent grinding for five to seven minutes occupies “no man’s land”: too slow for conversation, too short for substantial work. Blitzy instead runs overnight or all week and returns a large pull request.
6. An 86.8% SWE-bench result nearly saturates the current yardstick
Wissner-Gross explains SWE-bench as a measure of whether an AI system can resolve typical GitHub issues: understand a repository, fix a bug or performance problem, submit a pull request, and satisfy tests. SWE-bench Verified narrows this to 500 problems vetted as solvable and worthwhile by OpenAI researchers.
Pardeshi says the filmed leaderboard topped out at 75.2%; Blitzy scored 86.8% after verification with the SWE-bench CLI. Although the benchmark uses 12 repositories, its 500 branches represented roughly 400 million lines when ingested on Blitzy, within the one-to-two-billion-line cumulative corpus Blitzy estimates depending on whether updates are counted.
Reproducibility is central to the announcement. Pardeshi says Blitzy added no special SWE-bench scaffolding or benchmark-only features, contrasting that with reports in which a frontier model advertised near 80% but reproduced around 60%. “We care deeply about reproducibility and the practical real-world applicability.”
Diamandis challenges whether an 80%-plus benchmark is already saturated. Wissner-Gross argues that at 86.8%, much of the residue may be flawed rather than harder. He wants holdout benchmarks involving Linux at roughly 20 million lines and VS Code at roughly 4 million; the closing discussion says Blitzy is working with MIT on a successor.
7. Orchestration turns every frontier-model release into a product upgrade
Elliott’s answer to the Mag 7 threat is that they already sit inside Blitzy. The company uses Gemini, Anthropic, and OpenAI models against one another; hundreds of combinations of models, prompts, and tools raise quality beyond what a single model can deliver.
Blitzy calls the mechanism “extended inference-time validation,” paired with domain-specific context engineering. At every stage, the agents must retain the correct functional context despite an enterprise-scale repository, validate candidate output, and continue iterating rather than treat the first generated code as finished work.
Pardeshi says they made this bet when models had 5,000-token context windows: context would expand and coding would improve, so competing with the Mag 7 by training another model was unnecessary. They could instead “stand on the shoulders of giants”; Diamandis frames the frontier labs’ investment as a trillion dollars of R&D that Blitzy can ride, while Elliott says Blitzy improves whenever the underlying models improve.
8. The “great refactor” is technically plausible before it is commercially funded
Wissner-Gross proposes turning Blitzy loose on the “palimpsest” of Linux, Python, GNU, and other dependencies supporting modern civilization. His favored grand project is rewriting vulnerable legacy libraries in Rust or another memory-safe language: remove whole classes of weaknesses by rebuilding the software supply chain on which the stack depends.
Pardeshi says early experiments already include converting a 20-year-old, Windows-specific MATLAB library to Python and making it OS-agnostic. In another test, Blitzy selected an open issue in an NVIDIA repository and returned a pull request that solved the issue.
A sharper commercial demonstration used an AWS repository intentionally designed as messy mainframe code written in conflicting historical styles. Blitzy moved it from COBOL to Java autonomously, with the ability to compile out of the box; runtime work remained, but the inference took about a week against a project Elliott characterizes as normally multi-year.
Pardeshi draws the moat precisely: “Getting code from AI is a commodity.” Value appears when the rewrite must preserve existing behavior, compile, pass every unit test, avoid new security vulnerabilities, and satisfy enterprise-specific constraints. Without those validations, a chatbot-generated refactor is not a deployable asset.
9. Enterprise economics favor validated output, not the cheapest line of code
Elliott says a Blitzy line of code costs roughly 100 times one from another provider, but perhaps 100 times less than a human-written line. Quality is purchased deliberately because cheap code that triggers expensive review, diagnosis, and rework transfers cost rather than eliminating it.
Wissner-Gross reframes that premium through AI “hyperdeflation”: if inference costs fall by roughly 10x annually, a 100x premium resembles two years of cost decline. Elliott agrees that a civilization-scale rewrite is valuable today, but the immediate commercial question is “who’s the payer?”—so banks and insurers come first.
Elliott pushes back on Wissner-Gross’s labor-to-zero framing. Demand for software may be almost infinite, while Blitzy automates about 80% of the work on an average large-scale problem, knows exactly what it cannot finish, and hands a clean pull request plus an explicit task list to human developers.
The production claim is therefore 5x velocity, not a hypothetical 1,000x code-generation speed. Enterprises begin the next development sprint early; developers receive mostly written code and a guide to the remaining work. Elliott emphasizes that coordination, requirements, documentation, and release overhead determine realized productivity far more than token output alone.
10. The source of truth moves from code toward functional intent
Diamandis observes that legacy software treated working code as truth and documentation as a peripheral aid. If code can be regenerated overnight, the specification becomes the core asset. Pardeshi responds that the specification is still only an abstraction layer; Wissner-Gross adds that the human-readable specification is an intermediate representation over the system’s deeper functional understanding.
Blitzy’s deeper representation is a customer-specific hybrid graph-vector database capturing what the software must do independent of implementation language. That functional model, owned by the enterprise, can generate a human-readable specification or support migration from one language to another without losing required behavior.
Recursive self-improvement already has limits. Much of Blitzy’s sprint work is generated through Blitzy, but Elliott distinguishes replaceable software from core inventions—such as Pardeshi’s algorithms designed to ensure compilation and prevent circular dependencies. Rewriting the corpus may reproduce its shape without reproducing the underlying invention.
On METR-style autonomy horizons, Pardeshi says deployment, CI/CD, debugging, security analysis, monitoring, and tracing already exist separately; MCP and A2A can connect the components. He predicts end-to-end autonomous delivery is “months out” for projects meeting defined conditions, supported by product-manager, architect, QA, and prompt-writing agents already running in production.
11. Founders can compete by becoming indispensable customers of the giants
Elliott expects government modernization to become a major market: agencies possess vast legacy estates affecting identification, travel, and critical services. Blitzy does not yet serve the U.S. government, but he says “12 months” is plausible, while Diamandis argues that one successful agency deployment could cascade across the rest.
His strategic rule is mutual dependence: be a major customer of frontier labs so “you’re happy when they’re successful and they’re happy when you’re successful.” Blundin adds that large model companies benefit from thriving partner ecosystems, revenue, competition, and avoiding antitrust breakup; the startup should remain close enough to understand their direction.
Elliott’s protection against cloning is domain depth. Founders who have personally experienced enterprise security barriers, process gaps, and product failures can move faster than capital-rich incumbents burdened by bureaucracy. The minimal recipe is “the right talent, the right amount of capital,” and a problem validated with actual enterprises.
Elliott connects Blitzy’s intensity to his 2017 Ranger mission: roughly 100 people were sent covertly into Syria against about 2,000 ISIS fighters in Raqqa and had to recruit locally to retake the city. Blitzy screens first for ambition and invention, jokes that it is a “997” culture, and frames software automation as GDP expansion—Wissner-Gross’s closing imperative is simply: “Solve everything.”