🔬 Deep Research — astro.gpbot.net

How the Philippines Is Dealing with AI — Threat & Opportunity
2026-08-26
The Philippines faces AI as both its biggest threat and opportunity. Threat: the BPO/IT-BPM industry (~1.9M workers, ~$40B, ~8-10% of GDP) is highly AI-exposed — IMF estimates ~14% of the workforce at direct displacement risk, World Bank 35-37% of jobs at risk, ILO says 12.7M Filipinos (>1 in 4) exposed to GenAI, the highest in SE Asia. Opportunity: AI could add ~PHP 2.8T (~$50B) to GDP by 2030; a data-centre boom (~500MW now, ~1-1.5GW by 2028); NAISR 2.0 strategy. Institutional consensus (World Bank WDR 2026): emerging economies have 'more to gain, less to fear' — but must build connectivity, compute, skills, and pivot the BPO from volume to value.

⚠️ The AI threat (job displacement & BPO exposure)

  • BPO/IT-BPM is the core exposure: ~1.82M FTE jobs, ~$38-40B revenue (2024-25), ~8-10% of GDP — the world's #2 outsourcing destination behind India.
  • IMF: ~1/3 of PH occupations highly AI-exposed; ~14% of workforce at direct displacement risk, 22% to see major task changes. World Bank: 35-37% of jobs at risk.
  • ILO: 12.7M Filipinos (>1 in 4 workers) exposed to generative AI — the HIGHEST in Southeast Asia.
  • Avasant (2024): up to 89% of the BPO workforce at high automation risk.
  • Most at-risk jobs: clerical support, service/sales, technicians — concentrated among young, urban, female, college-educated service workers.
  • Oxford Economics/Cisco: ~1.1M PH jobs could be displaced by automation by 2028 (offset by new-demand job creation).

💼 The BPO transformation (the live case study)

  • The 2028 downgrade: IBPAP cut its targets from $59B/2.5M workers (2022 Roadmap) to $43.3-50.5B and 1.85-2.14M workers (mid-2026) — now 'AI-enabled'.
  • Adoption: ~67% of PH outsourcing firms are implementing AI; ~80% investing in upskilling toward higher-value roles.
  • No mass layoffs YET: affected entry-level workers are being redeployed; the shift is to higher-value 'AI-enabled' services rather than headcount cuts.
  • Pivot: 'from capacity to capability, volume to value' — building a pool of 2M AI-enabled 'Digital Filipino Workers'.
  • Competition: facing pressure from South Africa, Egypt, Poland, Colombia, Costa Rica, and Vietnam — as India moves up to software/tech while PH holds voice/customer service.
  • Real stories (BBC, Aug 2026): documented workers who 'trained the AI that replaced them'.

📈 The AI opportunity

  • Economic upside: AI could add PHP 2.8T (~$50.7B) to PH GDP by 2030 (Access Partnership/Google); another estimate ~PHP 1.3-1.8T (~7% of GDP).
  • Data-centre boom: ~500MW across ~28 colocations (2026) → market tripling from $735M (2025) to ~$2.5B by 2031; capacity forecast to exceed 1GW by 2028, 1.5GW online by 2027; DICT plans 18GW over a decade.
  • Government e-services: eGovPH super-app — 800M+ transactions, 56M downloads; Google Gemini Enterprise rolling to 50,000-200,000 public officers.
  • Talent: 800k+ graduates/yr, 200k+ STEM; TESDA free AI courses; Google AI Essentials; DICT SPARK/digitaljobsPH.
  • Advantages: young English-speaking workforce, US-friendly time zone, established IT-BPM base, improving subsea cables.
  • Foreign investment: UAE 4,000-acre AI-native Luzon hub; Microsoft $50B 'Global South' pledge (PH-specific $4B unverified); Google AI Opportunity Fund ($2.5T APAC benefit by 2030).

🏛️ Government policy response

  • NAISR 2.0 (National AI Strategy Roadmap, launched July 2024, DTI) — 2 pillars, 4 dimensions, 7 imperatives; establishes the Centre for AI Research (CAIR).
  • DOST AI National Strategy (2024-2028) — 5 pillars; aims to make PH an AI hub by 2040; 26-fold HPC increase by 2028; National HPC Centre.
  • DICT: 'Digital Bayanihan' strategy; committed to upskilling 300,000+ outsourcing workers; rejects a broad new AI law.
  • NEDA: policy note proposing an AI institute to skill workers; ADB designated PH for AI-readiness technical assistance (July 2026).
  • CREATE MORE Act: 200% enhanced deduction on power expenses + tax/duty-free capital equipment (key for high-cost PH electricity).
  • Legislation: House Bill 7396 (AI Development and Regulation Act) pending.

🌏 The emerging-economy view (IMF / World Bank / ADB)

  • IMF: AI exposure ~60% of jobs in advanced economies, ~40% in emerging markets, ~26% in low-income — EMs face less immediate disruption but lack infrastructure/skills to capture gains.
  • World Bank WDR 2026 ('The Promise of AI'): 'Developing economies today have more to gain — and less to fear — from AI than richer economies.' AI as a compressed catch-up tool; middle-income countries already account for half of ChatGPT's traffic.
  • IFC: emerging-market data-centre investment rose from $6B (2015) to $31B (2024); AI FDI hit $320B in 2025 (~40% going to EMs).
  • ADB: assessing member AI readiness; designated the Philippines for technical assistance to boost AI readiness.
  • The 4Cs framework (World Bank): connectivity, compute, context, competency — the foundation for AI gains.

🎯 Bottom line

  • Net job creator, but slower and more skills-intensive: BPO revenue keeps growing (~$42B 2026) and jobs rise (to ~2M), but ambitions were slashed ~25%.
  • The risk is less mass unemployment than reduced entry-level absorption and wage polarization — unless upskilling keeps pace.
  • Second-tier in the SE Asia AI race: Singapore leads (~75% of regional AI VC); PH competes on talent cost, English, and BPO scale, not compute or VC.
  • The core dilemma: PH's flagship industry is exactly what AI disrupts first — the 'first rung' of its labour market — so the pivot from volume to AI-enabled value is existential, not optional.
  • Institutional verdict: more to gain than fear, IF PH builds connectivity, compute, skills, and data foundations.

🏛️ Philippines rated on 'Why Nations Fail' institutional criteria

  • Inclusive Economic Institutions:
  • | Criterion | Rating | Key evidence |
  • |---|---|---|
  • | Secure Property Rights | ⚠️ Weak | Property rights 45.8/100, judicial effectiveness 41.8/100 (Heritage 2026); weak contract enforcement, titling gaps |
  • | Level Playing Field | ⚠️ Weak | Oligopolistic markets (banking, utilities, media), economic dynasties, ~106 days to register a foreign firm (bottom quintile); young, lightly-enforced competition commission |
  • | Encouraged Innovation | ✅ Moderate | Top 50 in Global Innovation Index 2025 (rank 50/139), #1 worldwide in high-tech exports — but import/assembly-driven, weak domestic R&D |
  • | Accessible Education | ⚠️ Weak | PISA 2022 near-bottom (math 355, reading 347); ~90% of 10-year-olds in 'learning poverty' (World Bank/UNICEF) |
  • | Creative Destruction | ✅ Moderate | MSMEs = 99.5% of businesses, ~63% of jobs, ~40% of GDP — dynamic small-firm churn, but concentrated incumbency blunts it |
  • | Public Services / Infrastructure | ⚠️ Weak | Recurring power-grid yellow/red alerts, archipelagic transport deficit; ~52nd/67 in World Competitiveness Yearbook 2024 |
  • Inclusive Political Institutions:
  • | Criterion | Rating | Key evidence |
  • |---|---|---|
  • | Political Pluralism | ✅ Moderate | Freedom House 58/100 'Partly Free'; competitive elections, lively press — but ~65-80% of elective posts held by political dynasties |
  • | Centralised State | ✅ Moderate | Strong unitary executive, but weak nationwide capacity — insurgency in Mindanao, uneven service delivery, low political-stability scores |
  • | Rule of Law | ⚠️ Weak | CPI 2025: 32/100 (rank 120/182); WGI Rule of Law ~47th percentile; slow, politicised courts, impunity |
  • | Public Accountability | ⚠️ Weak | WGI Control of Corruption ~36th percentile; pervasive, largely unpunished corruption; weak Ombudsman/COA enforcement |
  • 🎯 Overall verdict: The Philippines is a mixed but institutionally EXTRACTIVE-LEANING case. It has strong inclusive impulses — an entrepreneurial MSME-driven private sector, top-tier high-tech trade, improving innovation, and pluralist politics. But it scores weak on 7 of 10 pillars, crucially the ones Acemoglu & Robinson argue are the backbone of inclusive growth: property rights, rule of law, corruption control, education, and public services. Dynastic political capture, conglomerate oligopoly, weak courts, and chronic corruption are classic extractive-elite structures. On the WNF continuum, it is closer to extractive than inclusive — 'partially inclusive,' where institutional deficits (not a shortage of private dynamism) cap development.
Sources: IMF, World Bank WDR 2026, IFC, ADB, ILO, OECD.AI, UNESCO, IBPAP, NEDA/DTI/DICT/DOST, Reuters, BBC, Inquirer, BusinessWorld, Access Partnership/Google, White & Case, Kearney
🔗 World Bank WDR 2026: The Promise of AI · IMF: AI exposure by economy · OECD.AI: Philippines NAISR 2.0 · Inquirer: PH BPOs slash growth goals on AI shift · IFC: Accelerating AI investment in emerging markets
OpenAI's 'Jalapeño' Custom AI Inference Chip
2026-08-26
OpenAI's first in-house AI accelerator — the 'Jalapeño' Intelligence Processor, an LLM-inference ASIC co-designed with Broadcom, manufactured by TSMC, unveiled June 24 2026. It's inference-only (OpenAI keeps buying NVIDIA/AMD for training). First results (Aug 25 2026) claim 1.5-1.9x more AI work per watt and 1.7-3.6x lower latency than GB200/GB300. SemiAnalysis says it beats every NVIDIA/AMD/Google chip on inference efficiency tested. Captive to OpenAI; ramps at gigawatt scale from end-2026.

🌶️ What is Jalapeño?

  • OpenAI's first custom AI chip — an 'Intelligence Processor', a purpose-built ASIC for LLM INFERENCE (not training, not a GPU).
  • Partners: co-designed with Broadcom (silicon, packaging, Tomahawk networking); boards/racks by Celestica; fabricated by TSMC.
  • Captive silicon — built for OpenAI only, not sold externally.
  • Unveiled June 24, 2026 (Hot Chips); first engineering samples run GPT-5.3-Codex-Spark at production speed.
  • 9-month design-to-tapeout using AI-assisted design (OpenAI's claim).

⚙️ Technical specs (official vs SemiAnalysis)

  • Official (confirmed): inference ASIC; TSMC fab; Ethernet networking (Broadcom Tomahawk); 700 W package TDP (measured ≤550 W); deployment end-2026 at gigawatt scale.
  • SemiAnalysis (leaked/speculative): TSMC N3P reticle-sized compute die (~800mm²) + N3E I/O chiplet, 2.5D CoWoS package; HBM4 at 15.4 TB/s (highest per-package bandwidth); ~6 HBM stacks.
  • Interconnect: custom scale-up fabric (NVLink competitor, NOT UALink) on Tomahawk 6; local 128 ASICs/rack all-to-all; global 2,048 XPUs across 16 racks; PCIe Gen5 to host.
  • B0 stepping: 13.4 PFLOPs MXFP4 (vs 17.5 dense NVFP4 for a similar Rubin die, at much lower power).
  • Custom IP: MXFP matrix engine (weight-stationary systolic array, TPU-like), 64-bit scalar + FP32/INT32 vector cores, out-of-order cores with L1 cache; software via Gluon kernels + 'Teacup' serving engine.

📊 Performance claims (Aug 25, 2026)

  • 1.5-1.9x more AI work per watt than GB200/GB300 at peak throughput (OpenAI, InferenceX benchmark on A0 silicon).
  • 1.7-3.6x lower end-to-end latency; 2.1-4.1x higher performance on interactive workloads (GPT-OSS 120B, DeepSeek R1 670B, Kimi K2.5 1T).
  • SemiAnalysis (independent): >700 tok/s/user at concurrency 1 (DeepSeek R1); ~1,400 tok/s/user (Kimi/GPT-OSS) — beats published July Vera Rubin output-throughput/MW.
  • Caveat: results are self-reported / vendor claims; only short-context workloads tested (AgentX long-context not yet run). TCO roughly competitive with NVIDIA Rubin.

🤖 Agentic work performance (the honest caveat)

  • ⚠️ Key caveat: no true agentic benchmark has been run. Jalapeño's published results (Aug 25 2026) are ALL single-turn, 8k-input/1k-output (8k1k) fixed runs on SemiAnalysis' InferenceX. The long-context multi-turn agentic benchmark (AgentX) has NOT been run on Jalapeño — confirmed by SemiAnalysis and the live AgentX dashboard.
  • The 'interactive workload' gains (2.1-4.1x) — these map to minimum time-between-tokens (TBT) and peak per-user decode throughput at low concurrency, on open-weight models (GPT-OSS 120B, DeepSeek R1 670B, Kimi K2.5 1T), vs GB200/GB300:
  • | Model | Min TBT (Jalapeño vs NVIDIA) | Tok/s/user |
  • |---|---|---|
  • | GPT-OSS 120B | 0.69 vs 1.87 ms (2.7x) | 1,459 vs 535 |
  • | DeepSeek R1 | 1.43 vs 5.90 ms (4.1x) | 700 vs 169 |
  • | Kimi K2.5 1T | 1.44 vs 5.48 ms (3.8x) | 694 vs 182 |
  • Low-concurrency (1 user): DeepSeek R1 >700 tok/s/user; Kimi ~1,400; GPT-OSS ~1,400. On GPT-OSS, Jalapeño's iso-interactivity throughput/MW is >50x GB200's concurrency-1 point; on Kimi ~700 tok/s/user vs '9x the next best chip at 100 tok/s/user.'
  • Speculative decoding: Jalapeño runs single-token prediction ONLY — its wins are achieved WITHOUT speculative decoding, while GB200/GB300/Rubin used MTP/spec decoding. Speculative decoding gives competitors ~3-5x cost-per-token reduction; once implemented on Jalapeño, it should serve tokens even more cost-effectively. Current figures thus UNDERSTATE its upside.
  • AgentX (1M+ token, long-context, multi-turn agentic coding): released Aug 23 2026 by SemiAnalysis; live results cover Kimi K3, DeepSeek-V4-Pro, MiniMax-M3, Qwen3.5, GLM-5.3 on MI355X, GB300/GB200, B300/B200/H200/H100 — NO Jalapeño entry. SemiAnalysis: '8k1k is a much easier workload to tune for' and warns frameworks that perform well on 8k1k may perform worse on AgentX.
  • 1.2ms TPOT (Codex CLI) claim: confirmed as reported by SemiAnalysis, but it's an OpenAI DEMO (Codex CLI running internal 'Raiku'/5.3-Codex-Spark), not a published or independently verified benchmark.
  • Bottom line on agentic: Jalapeño is DESIGNED for agentic workloads (KV-cache locality, balanced prefill/decode, one large network domain, no disaggregation), and its low-latency single-turn numbers are impressive. But its actual long-context, multi-turn, tool-calling agentic performance is UNMEASURED as of Aug 26, 2026 — treat agentic claims as inferred, not proven.

🎯 Why OpenAI built it (strategy)

  • NVIDIA dependency & cost: NVIDIA has ~75% gross margins and captures ~85% of AI chip revenue; inference is 60-80% of AI GPU spend.
  • Supply/allocation: a second source not gated by NVIDIA allocation.
  • Vertical integration: owning chip, kernels, memory, networking, serving ('compute flywheel').
  • Expected ~50% lower cost per token than current NVIDIA GPUs (Broadcom CEO claim, not independently audited).
  • NVIDIA relationship: COMPLEMENTARY, not replacement — OpenAI still buys NVIDIA/AMD for training and most inference. OpenAI-NVIDIA 10 GW/$100B partnership continues.

⚔️ vs other custom chips

  • | Chip | Builder | Workload | Notes |
  • |---|---|---|---|
  • | Jalapeño | OpenAI+Broadcom | Inference | Captive |
  • | TPU (Ironwood/v7) | Google+Broadcom | Train+infer | Rentable on GCP |
  • | Trainium3 | AWS/Annapurna | Train+infer | Rentable; ~half NVIDIA cost |
  • | Meta MTIA | Meta+Broadcom | Recs+infer | Captive |
  • | Microsoft Maia 200 | MS in-house | Train+infer | Captive; delayed |

📅 Timeline & rollout

  • Jun 24, 2026: unveiled; A0 samples running GPT-5.3-Codex-Spark.
  • Aug 25, 2026: first benchmark results; B0 in fab (~25% perf/watt better).
  • End of 2026: initial deployment at gigawatt scale with Microsoft + partners; ~100 MW next deployment.
  • 2027: production ramps (Broadcom committed 1.3 GW of capacity for 2027); most output Q4 2027.
  • 2028: full-scale production; Gen 2 deep in development, Gen 3 forming; 10 GW roadmap through 2029.

🎯 Bottom line

  • Jalapeño is a major milestone — OpenAI's first custom silicon, and SemiAnalysis rates it as beating NVIDIA Blackwell/Rubin on inference efficiency and perf-per-watt.
  • But it's inference-only and captive — it absorbs a slice of OpenAI's inference load; NVIDIA stays the primary supplier for training and most inference.
  • The '50% cheaper' and benchmarks are vendor claims — credible (SemiAnalysis lab-tested) but not yet independently audited.
  • Structural signal: even the largest NVIDIA customer is building its own silicon — the hyperscaler custom-chip trend (TPU, Trainium, MTIA, Maia) now includes OpenAI.
Sources: OpenAI (Jun 24 2026 + Aug 25 2026), Broadcom, Reuters, CNBC, WSJ, SemiAnalysis, Bloomberg, TechCrunch
🔗 OpenAI: Jalapeño announcement (Jun 24) · OpenAI: Jalapeño first results (Aug 25) · Reuters: OpenAI unveils custom chip · SemiAnalysis: Jalapeño better than NVIDIA · CNBC: OpenAI & Broadcom reveal Jalapeño
Explainer & Fact-Check: Why the U.S. Propped Up Japan's Yen
2026-08-25
Clear explainer of the US-Japan currency intervention (Jul 31-Aug 3 2026) plus a fact-check of CFR expert Brad Setser's article. The article checks out as highly accurate: the US really did buy yen (selling euros) for the first time since 1998, confirmed by Treasury Secretary Bessent; the yen was at a 40-year low; Japan's economy is genuinely stronger than the yen's level implies. Only minor caveat: Japan's primary-surplus 'this year' claim was superseded by a July 2026 revision.

💱 What happened (the event)

  • Fri Jul 31, 2026: the US intervened in currency markets, selling euros from its reserves and buying yen — coordinated with Japan. USD/JPY fell from ~159 to ~157.
  • Mon Aug 3: Treasury Secretary Scott Bessent confirmed it (CNBC): the US bought yen 'alongside Japan to curb currency volatility and reduce risks to Asian markets.'
  • First support for the yen since 1998 (Asian financial crisis). The 2011 Fukushima action was the opposite — weakening the yen.
  • Context: the yen had hit a 40-year low (~163-164/USD) in July 2026. President Trump (Aug 2): 'We're always there for Japan.'

🤔 Why did the US do this?

  • Allies + strategy: Japan is a key US ally, and both Washington and Tokyo saw the yen as too weak.
  • Asian FX pressure: a weak yen pressures other Asian currencies (won, yuan, etc.) and makes it harder for China to allow a slow CNY appreciation.
  • US reindustrialization: currency markets were saying 'invest in high-surplus Asia' rather than 'invest in the US' — against Trump's reindustrialization goals.
  • Rare but legal: the US Treasury has broad legal authority (Gold Reserve Act 1934 / ESF) to intervene, but has rarely used it in recent decades.

🇯🇵 What Japan needs next (per Setser)

  • BOJ rate hikes: the Bank of Japan (rate 1.0%) has been slow; short-term rates are below US rates (3.50-3.75%) AND below Japan's inflation (~1.6-2.2% CPI), so real rates are negative.
  • A credible intervention threat to deter speculators.
  • More hedging: Japanese investors hold massive unhedged foreign assets (~$3.5T net external assets); a shift toward hedging would create structural yen demand.
  • Reality check: intervention without BOJ hikes is likely a 'pause, not a turn.'

🔎 Fact-check verdicts (CFR / Brad Setser article)

  • US intervened Jul 31 selling euros, buying yen — ✅ ACCURATE (CNBC, FX data).
  • Bessent confirmed Aug 3 — ✅ ACCURATE (CNBC, CFR).
  • First yen SUPPORT since 1998; 2011 was opposite — ✅ ACCURATE (WSJ, CNBC, BBC).
  • Yen at 40-year low in July 2026 — ✅ ACCURATE (Reuters, WSJ: ~162-164/USD).
  • Trump 'always there for Japan' Aug 2 — ✅ ACCURATE (CNBC, Reuters, Guardian).
  • Treasury has legal power but rarely intervenes — ✅ ACCURATE (Treasury ESF, CRS).
  • Inflation-adjusted yen back to 1960s lows — ✅ ACCURATE (FRED/BIS real EER ~62).
  • Japan current-account surplus ~3.8% of GDP — ✅ ACCURATE (IMF; record ¥17.4T H1).
  • Japan large external assets (~$3.5T, #3 creditor) — ✅ ACCURATE (Reuters; record ¥561.75T).
  • Net govt debt trending down vs GDP — ✅ ACCURATE (IMF 204% gross, declining; Fitch).
  • Headline fiscal deficit ~1% of GDP — ✅ ACCURATE (OECD ~-1.0% 2025; ~-0.6% TE).
  • Primary surplus 'this year' — ⚠️ TIME-SENSITIVE: FY2026 budget projected a first-in-28-years surplus (¥1.34T), but the July 30 2026 Cabinet Office revision pushed it to FY2027.
  • BOJ slow; rates below US and below Japan inflation — ✅ ACCURATE (BOJ 1.0% vs Fed 3.50-3.75%, CPI ~1.6-2.2%).

🎯 Bottom line

  • The CFR article is accurate, balanced, and from a top-tier source. Brad Setser (former US Treasury currency official) is highly credible on this exact topic.
  • 12 of 13 claims verified accurate; 1 is time-sensitive (Japan's primary-surplus timing). No claims are wrong or fabricated.
  • Japan's economy is stronger than the yen's level implies — big current-account surplus, record external assets, falling net debt.
  • The real fix is BOJ rate hikes — intervention alone is a pause, not a lasting turn.
Sources: CFR (Brad Setser), CNBC, WSJ, Reuters, Bloomberg, Guardian, BBC, Treasury ESF, CRS, FRED/BIS, IMF, OECD, Fitch, Cabinet Office
🔗 CFR article (source) · CNBC: Bessent confirms · Reuters: yen 40-year low · FRED: yen real effective rate · Reuters: Japan net external assets