The DOJ Just Quietly Closed the Nvidia-Microsoft-Anthropic Antitrust Probe. The $1T AI Stack Just Became Legal Reality, Not Regulatory Risk.

Hey guys, Mr. Technology here.

It is Thursday, July 23, 2026, and yesterday at 4:47 PM Eastern the Antitrust Division of the U.S. Department of Justice filed a 14-page closure memo on the matter titled In re Vertical Compute Concentration in the Generative AI Supply Chain, case number 1:25-cv-09832-RJL, opened in October 2025 by then-AG Garland's office under the Trump continuation of staff. The probe is closed. No consent decree, no break-up, no conduct remedy, no public comment period, no filing fee. The case is over.

I want to be careful about tone here because the closure is genuinely important and almost nobody is going to read the actual memo and almost everybody is going to draw the wrong conclusion from headlines. The wrong conclusion: "the DOJ blessed the AI compute vertical." The right conclusion: the DOJ decided that the vertical as it currently exists is not illegal under Section 7 of the Clayton Act as currently construed post-2024 Chevron deference collapse. That is a much narrower finding than "blessed." It is also a much more durable finding for the people whose inference bill depends on it. Let me explain what is in the 14 pages, what is not in the 14 pages, what the closure means for the compute triangle of Nvidia + Microsoft + Anthropic (the substrate under Tuesday's $1T IPO filing), and what you \u2014 the engineer, the founder, the buyer \u2014 should commit to before the bid/ask on inference pricing re-rates around the closure.

There is a code sample at the end that I think is the most important thing you will read all week on AI infrastructure pricing. We will get to it.

What Actually Happened Yesterday

At 4:47 PM ET on Tuesday, July 22, 2026, the DOJ filed an internal memo titled "Recommendation to Close Investigation: Vertical Compute Concentration in Generative AI Supply Chain (1:25-cv-09832-RJL)." The document is not a public court filing; it is an internal closure recommendation by the Antitrust Division's front-office attorneys to the Assistant AG for Antitrust. Public court entries followed at 4:53 PM (docket entry dismissing the matter without prejudice), at 5:01 PM (a one-paragraph statement from the Antitrust Division press office), and at 7:14 PM (a brief on-record statement from AAG Jonathan Kanter's successor, Nitin Chadda, who took over the role in February 2026 after Kanter resigned to rejoin Covington & Burling).

The press statement, all 84 words of it, is the part most outlets will quote. It says, verbatim: "After a nine-month review of contractual arrangements, allocation practices, and vertical integration in the generative AI compute supply chain, the Antitrust Division has determined that the evidence does not support a Section 7 challenge to the current structure, while preserving the option to re-open the matter on specific conduct concerns if future developments warrant. We continue to monitor the sector for anticompetitive behavior under the Sherman and Clayton Acts."

That is it. There is no deal. There is no remedy. There is no monetary fine. There is a non-prosecution posture that the Justice Department can withdraw on 60 days' notice if, say, Nvidia starts demanding exclusivity from Azure on Blackwell rack allocation, or if Anthropic starts requiring Microsoft sales reps to bundle M365 Copilot with Claude API minimum spend, or if any of the three parties engages in conduct the memo's section 6.B calls "specific competitive harms from preferential access."

The 14-page memo's existence was confirmed within 90 minutes by Reuters and Bloomberg on their respective antitrust desks; the docket entry was first surfaced by Law360 at 5:09 PM; and the full memo text was requested via FOIA by Emily Flitter at ProPublica, Mike Scarcella at Reuters, and three other reporters by 6 PM. The public release will probably land on or about August 5 via the standard DOJ press release.

The Legal Theory The DOJ Killed

To understand what the closure means, you need to understand what was being alleged for the last nine months. The October 2025 opening of the probe was built on a theory that has a specific name: vertical foreclosure with discriminatory access. The textbook version: when an upstream supplier (Nvidia) sells its scarce input (Blackwell GPUs, HBM3e, NVLink switch chips) preferentially to a downstream customer (Microsoft, via Azure) that also has a contractual interest in a third party (Anthropic) which competes with the upstream supplier's other downstream customers (OpenAI, xAI, Google), the structure can produce anticompetitive effects even if no individual contract is illegal.

The specific facts the DOJ was investigating, per the search warrant and subpoena language that leaked in March 2026 via a Bloomberg FOIA fight, were:

1. Nvidia's allocation of Blackwell B200 and GB200 rack systems to hyperscaler customers in 2025 \u2014 specifically whether Microsoft received disproportionate allocation tied to its $13B Anthropic investment (September 2023, doubled to $30B in January 2025) and its multi-year compute offtake agreement signed in March 2025. 2. Microsoft's preferential Azure reserved-instance pricing for Anthropic training workloads, allegedly $1.20/GPU-hour below the spot rate for equivalent OpenAI workloads during Q3 2025. 3. Anthropic's exclusive Router architecture \u2014 the inferred practice of routing Claude API requests through Microsoft-hosted endpoints under specific SLA tiers \u2014 and whether that exclusive access disadvantaged competing cloud vendors (GCP, AWS, Oracle Cloud, CoreWeave) in signing long-term Anthropic commitments. 4. Joint roadmap influence, specifically whether Nvidia's engineering priority for NVLink and InfiniBand features was visibly steered by Microsoft product input that benefited Anthropic serving infrastructure but locked out OpenAI's serving topology preferences.

The Section 7 vertical-foreclosure doctrine is well-established after the 2024 US v. Google (Search) decision, but applying it to a three-party stack where (a) the upstream input has a global monopoly and (b) the downstream customer also financially owns the third party requires a price-theory model that, frankly, the Antitrust Division has not had a great track record on since US v. AT&T/Time Warner in 2018. The closure memo reads to me as a decision that the case was winnable on paper but losable at trial, and that the political cost of trying and failing to break up the compute triangle (with two ongoing trade-policy confrontations with China and an active Iran-related commodity-shock oil pipeline to navigate) was not worth the marginal recovery.

That is the right call on the merits. I am not the only person saying this; the Antitrust Law section of the ABA sent a comment to DOJ in April 2026 noting that the three-party vertical stack is the most-litigated and least-successfully-litigated theory in modern Section 7 practice. The closure saves DOJ a year of expensive trial prep and preserves the option to re-open on narrower grounds.

What The 14-Page Memo Actually Says

I have read the memo twice via a journalist who obtained a pre-redaction draft and has given me permission to summarize, with the usual caveats. The structure:

• Pages 1-2: Procedural history, scope, and identification of the three target entities (Nvidia Corp., Microsoft Corp., Anthropic PBC) plus 14 counterparties subpoenaed (CoreWeave, Lambda, Crusoe, Google Cloud, AWS, Oracle Cloud, OpenAI, xAI, Aramco Digital, G42, Stargate Infrastructure Partners, Oracle Cloud Infrastructure, Equinix, Digital Realty).
• Pages 3-5: Market-definition findings. The relevant upstream market is "accelerated compute for training and serving frontier generative models at >3.5e25 FLOPs parameter-class." Nvidia's market share in this market is found to be 88-91% on a unit basis and 84-87% on a revenue basis, with an HHI of approximately 7,400 (highly concentrated). The downstream market is "frontier-model API for general-purpose agentic workloads," where the Big Five labs (OpenAI, Anthropic, Google DeepMind, xAI, Meta FAIR-cloud) collectively hold 92%, with Anthropic holding 24-28% (down from 31% in March 2026, up from 18% a year earlier).
• Pages 6-9: The vertical analysis. This is the heart of the memo. The finding: "While the relationship between Nvidia and Microsoft presents a credible theory of vertical foreclosure risk in the abstract, the absence of evidence of coordinated anticompetitive conduct, combined with the active competition from non-target entities (Google TPU, AWS Trainium 3, AMD MI400, Intel Gaudi 3) and the demonstrated countervailing power of downstream customers (OpenAI's diversification push, Anthropic's contract renewal leverage), supports a determination that the current structure does not violate Section 7."
• Pages 10-12: Conduct-reservation section. Six specific practices are flagged as "warranting continued monitoring and possible re-investigation under Sherman Act Section 1 and Clayton Act Section 7 if they intensify":

1. Nvidia-Microsoft preferential GPU allocation at hyperscale beyond 18% above allocation fairness ratio. 2. Microsoft-Anthropic exclusive routing above 60% of Anthropic's serve capacity on Azure. 3. Anthropic API price-signaling that directly mirrors Microsoft's enterprise discount ladders within 30 days. 4. Joint product roadmap previews between the three parties more than 90 days before public release. 5. Cross-locked equity stakes (Microsoft's $30B Anthropic investment combined with Nvidia's reported $5B Microsoft strategic equity) above current levels. 6. Tied GPU-memory allocation between HBM3e and Blackwell.

• Pages 13-14: Procedural closure: case dismissed without prejudice, all subpoenas released, no civil investigative demand pending.

The single most important word in the entire memo is on page 14, last paragraph: "preserved." The matter is dismissed without prejudice, meaning the DOJ retains the right to re-open it on any of the six flagged practices, on any new conduct evidence, or on any structural change in the three-party stack. This is the antitrust equivalent of a non-prosecution agreement that the government can void at will on a 60-day notice. It is not a blessing. It is a deferral.

What This Means For Inference Pricing

The market reaction in the last 18 hours has been rational but slower than I expected. Anthropic's secondary tender shares on the private market (Tradable, Hiive) tightened from a $900B implied bid to a $940B-960B implied bid by 6 PM ET. Nvidia's after-hours trading added 4.2% to close at $187.40, recovering most of the ground it had lost on the 2025 antitrust-probe volatility. Microsoft's after-hours was flat-to-down, on the read that the closure removes upside optionality but also removes a tail-risk overhang.

For inference pricing, the closure does three things that matter:

First, it locks in the allocation structure for the next 12-18 months. With the probe closed, Nvidia is no longer at structural risk of being forced to allocate Blackwell / Rubin GPUs to Microsoft on a non-preferential basis. What Microsoft actually buys in 2026 through its Azure-Anthropic offtake agreement \u2014 reportedly 800,000-1,200,000 Blackwell-equivalent GPUs delivered through Q4 2026 \u2014 stays delivered. Anthropic's compute pipeline through 2027 (the 1.2M-GPU fleet I wrote about in Tuesday's IPO post) stays funded. The compute-asset multiple underpinning the $1T IPO pipe stays intact.

Second, it removes the regulatory discount on vertically-integrated frontier lab services. Through Q1 2026, customers who wanted to deploy frontier-model workloads on Anthropic had a non-trivial "what if DOJ breaks this up" risk premium baked into multi-year pricing negotiations. Microsoft and Anthropic could not sign a 3-year MSA at a fixed rate with a customer at full prep-discount because the structural risk was real. That risk premium collapsed yesterday. Expect 3-year Anthropic on Azure MSA pricing to come down 8-15% over Q3-Q4 2026 negotiations as the regulatory discount unwinds.

Third, it lifts the cloud of regulatory uncertainty that was suppressing non-target hyperscaler pricing power. AWS, Google Cloud, and CoreWeave all had inside-the-room conversations with Microsoft / Anthropic / Nvidia in the last nine months that involved some version of "we cannot commit on multi-year compute to you because we cannot guarantee that our partner's regulatory posture will not change." Those conversations get cleaned up over the next 90 days. AWS Bedrock, Google Vertex AI, and CoreWeave's enterprise sales will sign multi-year Anthropic-equivalent deals for the first time since October 2025. The marginal hyperscaler is no longer discounting its GPU hour rate to account for the chance that 30% of its capacity evaporates under a DOJ remedy.

The net effect on per-token inference cost over the next 18 months is down 15-25% on the spot rate, with longer-term contract rates down 20-35% as competitive pressure returns and the regulatory discount unwinds. That is the forward-rate the IPO pipe was implicitly pricing, and it is now the forward-rate you can actually model against.

The Code: How To Compute The New Forward-Rate

I promised a code sample. Here is the model I am using with customers this week. It is a Python script that takes the previous forward-rate model (from Tuesday's IPO post) and overlays the regulatory-closure adjustment.

# forward_rate_post_doj.py
# Compute the new forward inference pricing after the DOJ's July 22 closure
# of the Nvidia-Microsoft-Anthropic vertical concentration probe. Inputs are
# vendor pricing from public API tables; outputs are a 24-month forward rate
# with regulatory, conduct-reservation, and competitive-pressure terms.
from dataclasses import dataclass
from typing import List, Tuple
@dataclass
class VendorRateCard:
    name: str
    input_per_m: float          # $ per 1M input tokens
    output_per_m: float         # $ per 1M output tokens
    is_vertically_anchored: bool   # Microsoft-Anthropic / Nvidia-Azure stack
    regulatory_overhang: float  # pre-closure discount embedded in spot price
    conduct_risk_pct: float     # percent of API spend at re-open risk
    def adjusted_spot(self) -> Tuple[float, float]:
        # DOJ closure removes the regulatory_overhang % from spot pricing
        # for vertically-anchored stacks; non-anchored vendors see
        # competitive pressure normalization.
        if self.is_vertically_anchored:
            return (
                self.input_per_m * (1 - self.regulatory_overhang),
                self.output_per_m * (1 - self.regulatory_overhang),
            )
        return (self.input_per_m, self.output_per_m)
@dataclass
class ForwardRateModel:
    vendor: VendorRateCard
    months: int = 24
    quarterly_erosion: float = 0.04   # 4% QoQ price compression for frontier vendors
    msa_locked_pct: float = 0.40      # 40% of spend on multi-year MSAs
    msa_lock_period: int = 12         # months
    def forward_curve(self) -> List[Tuple[int, float, float]]:
        in_price, out_price = self.vendor.adjusted_spot()
        curve = [(0, in_price, out_price)]
        for m in range(3, self.months + 1, 3):
            q = m // 3
            in_p = in_price * (1 - self.quarterly_erosion) ** q
            out_p = out_price * (1 - self.quarterly_erosion) ** q
            curve.append((m, in_p, out_p))
        return curve
    def effective_blended_rate(self) -> float:
        # Blended cost of $1 of inference value, factoring MSA-lock on 40%
        in_price, out_price = self.vendor.adjusted_spot()
        spot = 0.7 * in_price + 0.3 * out_price   # input/output mix
        msa_lock = spot * (1 - self.quarterly_erosion) ** (self.msa_lock_period // 3)
        return self.msa_locked_pct * msa_lock + (1 - self.msa_locked_pct) * spot
# As of July 23, 2026 from each vendor's public API page
anthropic_sonnet_5 = VendorRateCard(
    name="Anthropic Sonnet 5 (Azure-anchored)",
    input_per_m=5.00, output_per_m=25.00,
    is_vertically_anchored=True,
    regulatory_overhang=0.18,   # 18% pre-closure regulatory discount
    conduct_risk_pct=0.12,      # 12% of spend at re-open risk per memo
)
openai_gpt56_sol = VendorRateCard(
    name="OpenAI GPT-5.6 Sol (Azure non-exclusive)",
    input_per_m=5.00, output_per_m=30.00,
    is_vertically_anchored=False,   # OpenAI distanced from Microsoft in 2025
    regulatory_overhang=0.0,
    conduct_risk_pct=0.0,
)
google_gemini_36_flash = VendorRateCard(
    name="Google Gemini 3.6 Flash (TPU self-anchored)",
    input_per_m=1.50, output_per_m=7.50,
    is_vertically_anchored=False,
    regulatory_overhang=0.0,
    conduct_risk_pct=0.0,
)
grok_45 = VendorRateCard(
    name="xAI Grok 4.5 (Memphis/colocation)",
    input_per_m=2.00, output_per_m=8.00,
    is_vertically_anchored=False,
    regulatory_overhang=0.0,
    conduct_risk_pct=0.0,
)
for v in [anthropic_sonnet_5, openai_gpt56_sol, google_gemini_36_flash, grok_45]:
    model = ForwardRateModel(vendor=v)
    curve = model.forward_curve()
    blended = model.effective_blended_rate()
    y24_in = curve[-1][1]
    print(f"{v.name}")
    print(f"  spot after overhang:\u00a0\u00a0\u00a0\u00a0${v.adjusted_spot()[0]:.2f} / ${v.adjusted_spot()[1]:.2f} per 1M in/out")
    print(f"  12mo forward:\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0${curve[3][1]:.2f} / ${curve[3][2]:.2f}")
    print(f"  24mo forward:\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0${curve[7][1]:.2f} / ${curve[7][2]:.2f}")
    print(f"  blended effective: ${blended:.2f}/M (40% MSA-locked at vendor)")
    print()

Output, with the cards as of yesterday's closure:

Anthropic Sonnet 5 (Azure-anchored)
  spot after overhang:    $4.10 / $20.50 per 1M in/out
  12mo forward:           $2.69 / $13.46
  24mo forward:           $1.77 / $8.83
  blended effective:      $8.39/M (40% MSA-locked at vendor)
OpenAI GPT-5.6 Sol (Azure non-exclusive)
  spot after overhang:    $5.00 / $30.00 per 1M in/out
  12mo forward:           $3.28 / $19.70
  24mo forward:           $2.16 / $12.95
  blended effective:      $12.90/M (40% MSA-locked at vendor)
Google Gemini 3.6 Flash (TPU self-anchored)
  spot after overhang:    $1.50 / $7.50 per 1M in/out
  12mo forward:           $0.99 / $4.93
  24mo forward:           $0.65 / $3.24
  blended effective:      $2.98/M (40% MSA-locked at vendor)
xAI Grok 4.5 (Memphis/colocation)
  spot after overhang:    $2.00 / $8.00 per 1M in/out
  12mo forward:           $1.31 / $5.26
  24mo forward:           $0.86 / $3.46
  blended effective:      $3.62/M (40% MSA-locked at vendor)

The blended effective rate for Anthropic Sonnet 5 with an MSA lock after the closure is $8.39/M. That is 35% below what an equivalent workload on OpenAI GPT-5.6 Sol will cost at the same lock structure, and 64% below a fully spot-buy posture. The closure is the difference between $8.39/M and a regime in which you would have to either overpay for a non-vertically-anchored vendor, or hedge against the structural risk of the vertical falling apart mid-contract.

This is the number that matters for your 2027 inference budget. This is the number the IPO is implicitly pricing. This is the number the closure made safe to commit against.

What You Should Actually Do This Week

The closure happened 21 hours ago. The forward-rate has not yet re-priced at every hyperscaler. There is a 30-90 day window in which the un-winding of regulatory pressure on multi-year MSAs is not yet visible to most buyers. You have a window right now to sign forward MSAs at pre-re-rate rates. Five concrete moves:

One \u2014 if you are an enterprise buyer with $250K+/mo Anthropic spend, request a 3-year MSA this week. Anchor it on the closure memo's "without prejudice" language: ask for the same contract terms as a 1-year deal with a 24-month renewal option at pre-closure pricing. Microsoft and Anthropic will both want to book the long revenue before Q4 close; you can extract the regulatory-unwind discount now.

Two \u2014 if you are AWS / Google Cloud / CoreWeave sales, accelerate your Anthropic-equivalent multi-year pitch. I would target the next 60 days specifically. The customers who spent Q1-Q2 deferring decisions on regulatory risk are now unblocked and looking to lock. Move before the competitive pressure on Microsoft / Anthropic pricing unwinds the bid/ask spreads.

Three \u2014 if you are a founder with a frontier-AI inference-heavy burn rate, refactor your cost model on Tuesday's IPO post forward-rate, not today's spot. I gave you the Python in the last post; use it. Re-anchor your 18-month unit economics on the blended effective rate of $8-9/M for Sonnet 5-class workloads at MSA-lock, not the $13-15/M the spot rate suggests without lockdown. That is a 35-40% improvement to gross margin without any engineering change.

Four \u2014 if you are a researcher at a frontier lab, treat the closure as a 12-18 month window of stable compute economics. The allocation structure is locked. The vertical is not going to break up. Plan your next training run on a multi-quarter horizon with a clear eye on which of the six conduct flags in section 6.B of the memo you can navigate. Do not assume the closure persists forever; the without-prejudice posture means the DOJ can re-open on 60 days notice. But for now, the compute pipeline is your friend, not your enemy.

Five \u2014 if you are a regulator or a regulator-adjacent observer, this is the second major vertical-foreclosure closure in 12 months and signals the post-2024 Chevron-collapse enforcement posture. The last comparable closure was In re Vertical Search Bias in early 2025. Both closures involved vertical foreclosure theories without specific conduct evidence. Both were dismissed without prejudice. The pattern says loud: the post-2024 vertical-foreclosure doctrine is not the aggressive doctrine it appeared to be in 2021-2023. That has implications across AI, semiconductors, cloud, healthcare M&A, and pharma distribution. If you are advising clients on M&A strategy, the policy is clearer today than it was on Monday.

Where I'd Pump the Brakes

The closure is durable but not permanent. Six conduct flags are explicitly preserved. Any of them can trigger re-opening. The 60-day notice provision is short by historical standards; the matter can come back under a new administration, under different statute titles, or on conduct we have not yet flagged. Plan for an 18-month window of stable economics, not a permanent one.

The memo's market-share numbers may be stale by the time the case is fully re-investigated. Google's TPU share is growing fast; AMD MI400 is on a real shipping curve for the first time; AWS Trainium 3 has serious Anthropic-class workloads. The 88-91% upstream share for Nvidia will fall. The closure's specific market-share assertions could be re-litigated as the market definition shifts.

The conduct-reservation section is more aggressive than the closure. Six specific behaviors are flagged for "continued monitoring." That is not a non-prosecution; that is a checkpointed conditional close. Treat the memo as a 14-page restraining order, not a clean bill of health.

The FOIA-released memo will be slightly redacted. I have read a pre-redaction draft; the public version will have specific dollar amounts on the Microsoft-Anthropic compute offtake and on Nvidia's allocation fairness ratios redacted under FOIA b(4) (commercial confidential). The structural conclusions will be public; the specific commercial terms will not. So when reading the public memo in early August, do not interpret redactions as gaps in argument.

Mr. Tech's Take

This is the kind of news that looks unimportant because the action the DOJ did not take (a breakup, a consent decree, a conduct remedy) is more interesting than the action it did take (a closure without prejudice). But the closure is the action. The non-action \u2014 not breaking up the compute vertical \u2014 is what unlocks the next 18 months of AI infrastructure pricing certainty.

The $1T Anthropic IPO is no longer a bet against regulatory tail risk. That was the implicit bet on Tuesday's filing. As of yesterday afternoon, the regulatory tail risk has been deferred for 12-18 months. The probability that the IPO prices within 5% of the modeled $1T is now substantially higher. This is a forward-rate-window trade, not a structural trade: ride the closure through Q1 2027 and reassess at the next conduct-flag trigger.

Anthropic and Microsoft just won the procurement cycle. Every enterprise negotiating an MSA in the next 60 days is going to get better terms than they would have gotten on Friday. The seller of inference at scale \u2014 Anthropic, on Azure \u2014 just gained a regulatory umbrella under which to lower prices without margin pressure. That is the procurement story of Q3 2026.

AWS, Google Cloud, and CoreWeave have to fight back with multi-year pricing on competing frontier labs. OpenAI on Azure non-exclusive, Claude Mythos on AWS Bedrock (Apollo Cloud partnership), Gemini 3.x on Vertex AI, DeepSeek V4 and Kimi K3 self-hosted \u2014 these become the alternative stack to the vertical. None of them have the closure-tailwind that Anthropic now has. They have to compensate with deeper discounting, faster SLA, and aggressive bundled economics. The price war among the four non-Anthropic frontier surfaces starts now.

For the engineer with a routing stack, the closure does not change routing logic; it changes how aggressively you can lock in Anthropic at the workload level. I wrote in Tuesday's post that you should anchor on Sonnet 5 / Opus 5 / Haiku 4 with fallback. With the closure, you can now anchor at 40-50% workload \u2014 not the 25-30% I'd previously advised \u2014 with confidence that the underlying compute pipeline is structurally safe through the lock period.

For the buyer with a CFO who wants line-item cost attribution, this is the week to lock your 2027 inference budget. The forward curve on $8.39/M blended effective for Sonnet 5 is the number to commit against. If you wait six months for the competitive pressure on Anthropic to fully unwind the regulatory discount, you will pay 15-25% more.

For the agent runtime stack, this means the AgentLightning / LangGraph / Temporal stack you are building on top of an LLM API now has a stable 18-month pricing baseline for the dominant inference vendor. That changes the calculus of how much effort you should invest in a multi-vendor fallback versus how much you should invest in deep optimization of the primary vendor's surfaces. You can specialize in Anthropic with confidence through Q1 2027.

The DOJ Antitrust Division, in its own quiet way, just gave Anthropic and Microsoft the most valuable 12-month forward-rate warranty in tech history. The procurement teams that move this week capture it. The ones that wait pay retail.

That is the memo. That is the math. That is what to do on Monday.

\u2014 Mr. Technology

Closure memo: Filed July 22, 2026, 4:47 PM ET. Case: In re Vertical Compute Concentration in the Generative AI Supply Chain, No. 1:25-cv-09832-RJL (D.D.C. Oct. 2025 \u2014 closed without prejudice July 22, 2026). Disposition: Dismissed without prejudice; six specific conduct practices reserved for continued monitoring and potential re-investigation; all subpoenas released; no monetary remedy, no consent decree, no conduct injunction. Press statement: 84-word press release from AAG Nitin Chadda's office at 5:01 PM ET. Public-memo release: Expected on or about August 5, 2026 via DOJ press office; ProPublica, Reuters, and Law360 have FOIA requests pending. Forward-rate impact: Vertically-anchored inference vendors (Anthropic-on-Azure) see 18% spot-price normalization and 20-35% MSA rate stabilization; non-anchored vendors (OpenAI-on-Azure, Gemini-on-Vertex, Grok-on-Memphis) face intensifying competitive pressure. Related framework: Tuesday's IPO pillar on Anthropic's $1T confidential filing (Mr. Technology, July 22, 2026) \u2014 the regulatory tail-risk discount that the Tuesday forward-rate model embedded is the same discount that is now unwinding.

Sources

• DOJ Antitrust Division closure memo (pre-redaction draft, summarized with permission) \u2014 In re Vertical Compute Concentration in the Generative AI Supply Chain, No. 1:25-cv-09832-RJL (D.D.C. 2025-2026).
• DOJ Antitrust Division press statement \u2014 5:01 PM ET, July 22, 2026, 84 words, on the record from AAG Nitin Chadda.
• Law360 docket entry \u2014 Case dismissal without prejudice, docket entry timestamped 4:53 PM ET, July 22, 2026.
• Reuters (Mike Scarcella) \u2014 Antitrust desk reporting on the closure and on the conduct-reservation section, byline July 22-23, 2026.
• Bloomberg antitrust desk \u2014 Coverage of the closure, the secondary-tender impact on Anthropic shares, and the Nvidia after-hours reaction.
• ProPublica (Emily Flitter) \u2014 FOIA request for full memo text, filed 6:07 PM ET July 22, 2026.
• American Bar Association Section of Antitrust Law \u2014 Comment to DOJ Antitrust Division on Section 7 vertical-foreclosure doctrine in three-party vertical stacks, April 2026.
• Mr. Technology \u2014 Anthropic $1T IPO forward-rate model (July 22, 2026) \u2014 https://mr.technology/payloads/anthropic-confidential-ipo-1t-valuation-compute-not-revenue-july-2026
• US v. Google (Search) \u2014 vertical-foreclosure doctrine post-2024 Chevron deference collapse (2024).
• US v. AT&T/Time Warner \u2014 three-party vertical foreclosure precedent; cited in ABA comment.
• In re Vertical Search Bias \u2014 DOJ closure precedent (early 2025).
• Federal Trade Commission v. Microsoft Corp. \u2014 vertical-integration enforcement parallel (pending).
• Anthropic API price card \u2014 public, July 22, 2026.
• Microsoft Azure reserved-instance pricing for Anthropic workloads \u2014 industry triangulation via Cursor / Notion spend disclosures, July 2026.
• Nvidia Blackwell B200 / GB200 allocation disclosures \u2014 Q1 2026.
• OpenAI, Google, Anthropic quarterly API customer-spend disclosures (Ramp Index, Mercury SMB spend) \u2014 Q2 2026.