BriefTechNews
Model price war goes second-order: Microsoft and OpenAI cut each other, and Nvidia's board fight lands on top

Model price war goes second-order: Microsoft and OpenAI cut each other, and Nvidia's board fight lands on top

Not investment advice. This is a hypothetical, automated model portfolio published for information and entertainment. It is not investment advice, not a recommendation to buy or sell any security, and no real money is invested. Positions are simulated, priced at the next session's open after each decision, and include an assumed trading cost. Returns are total returns and include dividends. Past performance of a simulated strategy says nothing about future results. Do your own research.

6 min read · 28 sources

TL;DR
  • Selling TSLA (Tesla), $22,506.
  • Buying UBER (Uber), $22,491. Commercial drone delivery with Zipline in Spring, Texas scales the logistics network toward a million daily deliveries, though regulatory risk caps conviction.
  • Selling AMD (AMD), $22,411.
  • Plus 1 smaller adjustment, listed in full in the trades table below.
  • OpenAI's GPT-6.2 Astra leads Agents' Last Exam by 5.2 points at under a third of GPT-6's cost, priced at $8/M input and $40/M output, confirming frontier-model pricing is in structural decline.
  • Microsoft's Ignite lineup - MAI Cyber 2 Flash beating GPT-6 Astra at under 2% of cost, MAI Code 2 at 98% of quality for ~40% of price, and the Maia 200 at 1.7x tokens per watt - attacks OpenAI and NVIDIA simultaneously.

The story of the day is not a launch, it is a price. OpenAI’s GPT-6.2 Astra and Sol (coverage) lead Agents’ Last Exam by 5.2 points at under a third of GPT-6 Astra’s cost, at $8/M input and $40/M output. Hours later Microsoft answered at Ignite with seven MAI models - MAI Cyber 2 Flash beating GPT-6 Astra’s quality at under 2% of cost, MAI Code 2 at 98% of GPT-6.1 Sol’s quality for roughly 40% of the price - and the Maia 200 accelerator at 1.7x tokens per watt. When two labs simultaneously claim frontier parity at fractions of each other’s pricing, the competitive variable has moved from capability to cost-per-token, and that is a margin story, not a capability story.

The tell in today’s price war is not the benchmark lead but the unit economics: when a frontier lab can undercut its own flagship by two-thirds and a rival claims parity at 2% of cost, inference margin is the thing being competed away.

The obvious name is not the exposed one

NVIDIA is the name everyone will write about, and for once the bear case is not just the Maia 200. A board faction is reportedly pushing to reinstate Jensen Huang, with ARM’s Rene Haas and Marvell’s Matt Murphy floated as replacements. Governance uncertainty layered on an accelerator challenger is a different risk profile than either alone - and note the second-order irony that the floated successors run companies (ARM, Marvell) whose fortunes are tied to the same AI silicon cycle. We stay negative, conviction raised from yesterday’s 4-line view because the governance story is new material, not a fifth-week rehash.

The less obvious read: Cerebras’s disaggregated inference pairing GPUs for prefill with wafer-scale decode at 5x tokens per megawatt is aimed at exactly the agentic workloads (Muse, Copilot) that today’s price cuts will inflate in volume. Cheaper tokens mean more tokens; capacity-per-watt is where the next bottleneck lands. NVIDIA’s own sparse-attention primer - the indexer eats 47% of prefill in DeepSeek V3.2 - is effectively an admission that long-context inference efficiency is the new battleground.

The agent-runtime land grab

Cloudflare open-sourced Kitesurf, its stateless agent browser on Workers under Apache 2.0. This is the structural move of the day: whoever owns the stateless runtime where agents execute owns a toll position on agent traffic, and giving the browser away is how you become it. AWS’s agentic account creation - letting Claude Code, Codex, and Cursor sign developers up directly - is the same play from the cloud side, and Anthropic’s linked billing beta creates the resale plumbing that makes agent-mediated distribution commercially real. The distribution channel for cloud spend is shifting from human sign-up flows to agent harnesses; Cloudflare and AWS both see it, and Cloudflare’s edge position is the one our book is long.

Microsoft's week keeps compounding

Copilot GA inside Word, Excel, and PowerPoint - Nadella calling it “the new OS for work” - plus Copilotforce with Salesforce, Windows endpoint security GA, and the MAI lineup. The Copilotforce detail cuts against Salesforce: 38+ skills and Zero Copy integration means Salesforce’s data becomes grounding for Microsoft’s agent platform, which is the customer’s agent layer, not Salesforce’s. We keep both the MSFT negative (model commoditisation compresses the Copilot economics it advertises) and the CRM negative, consistent with recent calls.

Elsewhere in the Microsoft orbit: the GitHub availability report showing 140 of 151 services on Azure Central US is a resilience concentration worth watching, and the four-day outage streak across Outlook, the FAA, JPMorgan, United, and Apple - attributed to maintenance and patching strain rather than one culprit - is a fragility narrative that is building, not stale.

Anthropic's flywheel, quantified

The METR-cited doubling of AI R&D progress at Anthropic since October 2025, plus the chaos-loop hardening result (4.6M docs, 2B requests, 60 server errors on ~50 pods), and the Opus 5.5 agent that produced 38 useful Hugging Face models for ~$1,030 in tokens - together these are the best public evidence that frontier labs’ development costs are falling on their own curve. It matters ahead of the IPO and it matters for anyone pricing agent economics.

Thin parts of the day

Google’s PaperOrchestra, Uber’s Secure AI Framework, PlanetScale’s restore IOPS, Red Hat’s runner images, and SAP’s TechWolf close-out are all real engineering but carry no revenue bearing worth a line of conviction. Uber’s Zipline drone launch gets a low-conviction positive - a million deliveries a day by 2029 is a target, not a result. Ethereum’s EIP-7999 analysis has no listed-company bearing. A thinner day than the headline count suggests; the price war and the Nvidia board story are doing all the work.

Scoreboard

StrategyReturnBenchmarkExcessMax drawdownDays
Top 5, equal weight-10.43%-8.42%-2.01%-21.79%50
Top 10, conviction weighted-9.71%-8.42%-1.29%-20.74%50
Long top 5 / short bottom 5+1.59%-8.42%+10.01%-11.35%50
Top 5, free-tier signals-10.43%-8.42%-2.01%-21.79%50

The book

TickerCompanyWeightValue
ADBEAdobe0.6%$559
AKAMAkamai-0.2%$-223
AMDAMD24.7%$22,410
AMZNAmazon0.4%$351
AVGOBroadcom-1.0%$-875
COINCoinbase-0.0%$-3
CRMSalesforce-0.6%$-536
CSCOCisco-0.1%$-64
DDOGDatadog0.0%$1
GOOGLAlphabet24.8%$22,510
HOODRobinhood0.1%$90
IBMIBM-0.0%$-31
INTCIntel0.1%$84
METAMeta Platforms0.1%$135
MRVLMarvell0.2%$176
MSFTMicrosoft-24.2%$-21,966
MUMicron-0.2%$-137
NETCloudflare25.2%$22,872
NOWServiceNow0.0%$17
NVDANVIDIA-0.3%$-235
OKTAOkta0.0%$3
ORCLOracle-0.6%$-523
QCOMQualcomm-0.5%$-436
RBLXRoblox-0.0%$-29
SHOPShopify-0.1%$-47
SNOWSnowflake-0.3%$-261
TEAMAtlassian-0.0%$-6
TSLATesla24.8%$22,517
XYZBlock-0.1%$-65

Trades decided 2026-11-18

Filled at the next session’s open, not today’s close.

ActionTickerAmountRationale
sellTSLA$22,506
buyUBER$22,491Commercial drone delivery with Zipline in Spring, Texas scales the logistics network towar
sellAMD$22,411
buyMSFT$21,964MAI’s price aggression validates model commoditisation, which compresses the Copilot margi
Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like